Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 8 new columns ({'usable_memory_gb', 'bandwidth_gbs', 'tdp_w', 'memory_gb', 'msrp_usd', 'best_runtime', 'category', 'memory_type'}) and 17 missing columns ({'default_context_k', 'release', 'is_moe', 'pulls', 'commercial_use', 'q4_k_m_gb', 'params_b', 'q8_0_gb', 'min_ram_q4_gb', 'active_params_b', 'license_note', 'modality', 'family', 'ollama_tag', 'license', 'pulls_text', 'fp16_gb'}).
This happened while the json dataset builder was generating data using
hf://datasets/ansumanshah/local-ai-model-memory-requirements/devices.json (at revision a4cfe4710bea52eeadb5da63791165a4db0247bc), ['hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/audio-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/devices.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/image-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/tools.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/video-models.json'], ['hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/audio-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/devices.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/image-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/tools.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/video-models.json']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
id: string
name: string
category: string
memory_gb: int64
memory_type: string
usable_memory_gb: double
bandwidth_gbs: double
msrp_usd: int64
tdp_w: int64
best_runtime: string
sources: list<item: string>
child 0, item: string
notes: string
audio: null
-- schema metadata --
huggingface: '{"info": {"features": {"id": {"dtype": "string", "_type": "' + 656
to
{'id': Value('string'), 'name': Value('string'), 'family': Value('string'), 'params_b': Value('float64'), 'is_moe': Value('bool'), 'active_params_b': Value('null'), 'q4_k_m_gb': Value('null'), 'q8_0_gb': Value('null'), 'fp16_gb': Value('null'), 'min_ram_q4_gb': Value('null'), 'default_context_k': Value('null'), 'ollama_tag': Value('null'), 'release': Value('string'), 'pulls': Value('null'), 'pulls_text': Value('null'), 'modality': Value('string'), 'license': Value('string'), 'commercial_use': Value('string'), 'license_note': Value('string'), 'audio': {'subtype': Value('string'), 'backbone_params_b': Value('float64'), 'precision': Value('string'), 'recommended': Json(decode=True), 'cpu_ok': Value('bool'), 'device_classes': List(Value('string')), 'tools': List(Value('string')), 'task': Value('string')}, 'sources': List(Value('string')), 'notes': Value('string')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 8 new columns ({'usable_memory_gb', 'bandwidth_gbs', 'tdp_w', 'memory_gb', 'msrp_usd', 'best_runtime', 'category', 'memory_type'}) and 17 missing columns ({'default_context_k', 'release', 'is_moe', 'pulls', 'commercial_use', 'q4_k_m_gb', 'params_b', 'q8_0_gb', 'min_ram_q4_gb', 'active_params_b', 'license_note', 'modality', 'family', 'ollama_tag', 'license', 'pulls_text', 'fp16_gb'}).
This happened while the json dataset builder was generating data using
hf://datasets/ansumanshah/local-ai-model-memory-requirements/devices.json (at revision a4cfe4710bea52eeadb5da63791165a4db0247bc), ['hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/audio-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/devices.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/image-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/tools.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/video-models.json'], ['hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/audio-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/devices.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/image-models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/models.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/tools.json', 'hf://datasets/ansumanshah/local-ai-model-memory-requirements@a4cfe4710bea52eeadb5da63791165a4db0247bc/video-models.json']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
id string | name string | family string | params_b float64 | is_moe bool | active_params_b null | q4_k_m_gb null | q8_0_gb null | fp16_gb null | min_ram_q4_gb null | default_context_k null | ollama_tag null | release string | pulls null | pulls_text null | modality string | license string | commercial_use string | license_note string | audio dict | sources list | notes string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
whisper-large-v3 | Whisper large-v3 | whisper | 1.55 | false | null | null | null | null | null | null | null | 2023-11 | null | null | audio | MIT | yes | MIT per the OpenAI Whisper repo (the HuggingFace card labels large-v3 Apache-2.0; the repo is the source of truth). | {
"subtype": "stt",
"backbone_params_b": 1.55,
"precision": "int8 (faster-whisper) / fp16 (whisper.cpp)",
"recommended": "{\"gb\":2.5,\"quant\":\"int8\",\"source\":\"https:\\/\\/github.com\\/SYSTRAN\\/faster-whisper\\/issues\\/1030\",\"synthesis\":true}",
"cpu_ok": true,
"device_classes": [
"mac",
"... | [
"https://github.com/openai/whisper",
"https://huggingface.co/openai/whisper-large-v3",
"https://github.com/SYSTRAN/faster-whisper",
"https://github.com/ggml-org/whisper.cpp"
] | Most accurate Whisper (1.55B). Via faster-whisper int8 it peaks at ~2.5GB VRAM (2,953MB measured, beam=5); via whisper.cpp the ggml model is 2.9GB and runtime RAM ~3.9GB. Runs CPU-only and on Apple Silicon (Metal) and phones (whisper.cpp). The openai/whisper README's ~10GB figure is the original fp32 PyTorch path, not ... |
whisper-large-v3-turbo | Whisper large-v3-turbo | whisper | 0.809 | false | null | null | null | null | null | null | null | 2024-10 | null | null | audio | MIT | yes | null | {
"subtype": "stt",
"backbone_params_b": 0.809,
"precision": "int8 (faster-whisper) / fp16 (whisper.cpp)",
"recommended": "{\"gb\":1.5,\"quant\":\"int8\",\"source\":\"https:\\/\\/github.com\\/SYSTRAN\\/faster-whisper\\/issues\\/1030\"}",
"cpu_ok": true,
"device_classes": [
"mac",
"nvidia",
"amd"... | [
"https://huggingface.co/openai/whisper-large-v3-turbo",
"https://github.com/SYSTRAN/faster-whisper/issues/1030",
"https://github.com/ggml-org/whisper.cpp"
] | Pruned large-v3 (decoder layers cut from 32 to 4, ~809M params) that runs roughly 5x faster at near-large-v3 accuracy. Via faster-whisper int8 it peaks at ~1.5GB VRAM (1,545MB measured). Light enough for phones (CoreML) and CPU. License MIT. Sources: HF model card, faster-whisper benchmark, whisper.cpp README. |
whisper-small | Whisper small | whisper | 0.244 | false | null | null | null | null | null | null | null | 2022-09 | null | null | audio | MIT | yes | null | {
"subtype": "stt",
"backbone_params_b": 0.244,
"precision": "int8 / fp16",
"recommended": "{\"gb\":0.85,\"quant\":\"fp16 (whisper.cpp)\",\"source\":\"https:\\/\\/github.com\\/ggml-org\\/whisper.cpp\"}",
"cpu_ok": true,
"device_classes": [
"mac",
"nvidia",
"amd",
"laptop",
"iphone",
... | [
"https://github.com/openai/whisper",
"https://github.com/ggml-org/whisper.cpp"
] | The phone-friendly Whisper tier (244M). whisper.cpp runtime RAM is ~852MB; the ggml model is ~466MB. Runs comfortably CPU-only and on-device on modern phones. Good accuracy/speed balance when large-v3 is too heavy. License MIT. Sources: OpenAI Whisper repo, whisper.cpp README. |
kokoro-82m | Kokoro-82M | kokoro | 0.082 | false | null | null | null | null | null | null | null | 2025-01 | null | null | audio | Apache-2.0 | yes | null | {
"subtype": "tts",
"backbone_params_b": 0.082,
"precision": "fp32 / ONNX int8",
"recommended": "{\"gb\":1,\"quant\":\"fp32\",\"source\":\"https:\\/\\/huggingface.co\\/FluidInference\\/kokoro-82m-coreml\",\"synthesis\":true}",
"cpu_ok": true,
"device_classes": [
"mac",
"nvidia",
"amd",
"lapt... | [
"https://huggingface.co/hexgrad/Kokoro-82M",
"https://huggingface.co/FluidInference/kokoro-82m-coreml",
"https://github.com/remsky/Kokoro-FastAPI",
"https://github.com/puff-dayo/Kokoro-82M-Android"
] | Tiny 82M StyleTTS2-based voice model with surprisingly good quality. Weights ~326MB (fp32), under 1GB for CPU inference; CoreML process peak ~1.5GB on Apple Silicon; ~2-3GB on a CUDA GPU. Confirmed on-device iPhone/iPad (CoreML, Apple Neural Engine) and Android (ONNX int8). Apache-2.0, commercial OK. Sources: HF card, ... |
orpheus-3b | Orpheus 3B | orpheus | 3 | false | null | null | null | null | null | null | null | 2025-03 | null | null | audio | Apache-2.0 | yes | null | {
"subtype": "tts",
"backbone_params_b": 3,
"precision": "Q4_K_M GGUF (llama.cpp) / fp16 (vLLM)",
"recommended": "{\"gb\":4,\"quant\":\"Q4_K_M GGUF\",\"source\":\"https:\\/\\/huggingface.co\\/Mungert\\/orpheus-3b-0.1-ft-GGUF\",\"synthesis\":true}",
"cpu_ok": true,
"device_classes": [
"mac",
"nvidia"... | [
"https://huggingface.co/canopylabs/orpheus-3b-0.1-ft",
"https://github.com/canopyai/Orpheus-TTS",
"https://huggingface.co/Mungert/orpheus-3b-0.1-ft-GGUF"
] | A Llama-3.2-3B fine-tune that emits speech tokens, so it runs through the usual LLM stacks. The llama.cpp GGUF path (Q4_K_M ~2.5GB on disk) runs on CPU or Apple Silicon Metal at roughly ~4GB with overhead; the vLLM/CUDA path needs more VRAM headroom. Apache-2.0, commercial OK. Anchor is the Q4_K_M file plus typical lla... |
bark | Bark | bark | 0.9 | false | null | null | null | null | null | null | null | 2023-04 | null | null | audio | MIT | yes | null | {
"subtype": "tts",
"backbone_params_b": 0.9,
"precision": "fp32 / fp16",
"recommended": "{\"gb\":5,\"quant\":\"fp32\",\"source\":\"https:\\/\\/huggingface.co\\/blog\\/optimizing-bark\"}",
"cpu_ok": true,
"device_classes": [
"mac",
"nvidia",
"amd",
"laptop"
],
"tools": [
"HF Transfor... | [
"https://huggingface.co/suno/bark",
"https://huggingface.co/blog/optimizing-bark",
"https://github.com/suno-ai/bark"
] | Suno's 3-stage generative audio model (~300M per stage, ~0.9B total) that can produce speech, music and sound effects. Peaks at ~5GB VRAM in fp32 via HF Transformers; drops to ~1GB with fp16 + CPU offload, and bark-small needs ~1.9GB. CPU-only works but needs ~8GB RAM and is slow. MIT, commercial OK. Sources: HF model ... |
dia-1-6b | Dia 1.6B | dia | 1.6 | false | null | null | null | null | null | null | null | 2025-04 | null | null | audio | Apache-2.0 | yes | null | {
"subtype": "tts",
"backbone_params_b": 1.6,
"precision": "fp16",
"recommended": "{\"gb\":10,\"quant\":\"fp16\",\"source\":\"https:\\/\\/huggingface.co\\/nari-labs\\/Dia-1.6B\"}",
"cpu_ok": false,
"device_classes": [
"nvidia"
],
"tools": [
"PyTorch (CUDA)",
"nari-labs/dia"
],
"task": "T... | [
"https://huggingface.co/nari-labs/Dia-1.6B",
"https://github.com/nari-labs/dia/issues/34"
] | Nari Labs' dialogue TTS (1.6B). The model card states ~10GB VRAM peak (steady ~7.4GB); an RTX 3060 12GB user reported OOM at peak. CPU support is listed as not-yet-added, and no Apple Silicon or AMD path is confirmed, so it is NVIDIA-only for now. Apache-2.0, commercial OK. Sources: HF model card, GitHub OOM report. |
musicgen-small | MusicGen small | musicgen | 0.3 | false | null | null | null | null | null | null | null | 2023-06 | null | null | audio | CC-BY-NC-4.0 | no | Model weights are CC-BY-NC 4.0 (non-commercial); the AudioCraft code is MIT. | {
"subtype": "music",
"backbone_params_b": 0.3,
"precision": "fp32",
"recommended": "{\"gb\":3,\"quant\":\"fp32\",\"source\":\"https:\\/\\/github.com\\/facebookresearch\\/audiocraft\\/blob\\/main\\/docs\\/MUSICGEN.md\",\"synthesis\":true}",
"cpu_ok": false,
"device_classes": [
"mac",
"nvidia",
"... | [
"https://huggingface.co/facebook/musicgen-small",
"https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
] | Smallest MusicGen (300M), the most accessible music model. AudioCraft requires a GPU; the small model fits in roughly ~2-4GB VRAM (synthesis from the 300M fp32 weights plus EnCodec, no official figure). Weights are CC-BY-NC (non-commercial). Sources: HF card, AudioCraft MusicGen docs. |
musicgen-medium | MusicGen medium | musicgen | 1.5 | false | null | null | null | null | null | null | null | 2023-06 | null | null | audio | CC-BY-NC-4.0 | no | Model weights are CC-BY-NC 4.0 (non-commercial); the AudioCraft code is MIT. | {
"subtype": "music",
"backbone_params_b": 1.5,
"precision": "fp32",
"recommended": "{\"gb\":14,\"quant\":\"fp32\",\"source\":\"https:\\/\\/github.com\\/facebookresearch\\/audiocraft\\/blob\\/main\\/docs\\/MUSICGEN.md\",\"synthesis\":true}",
"cpu_ok": false,
"device_classes": [
"nvidia",
"amd"
],
... | [
"https://huggingface.co/facebook/musicgen-medium",
"https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
] | MusicGen medium (1.5B). AudioCraft's docs state it requires a GPU with at least 16GB of memory; peak consumed is ~14GB fp32 (synthesis from the official 16GB-GPU guidance and the 1.5B weight size). Weights are CC-BY-NC (non-commercial). Sources: HF card, AudioCraft MusicGen docs. |
musicgen-large | MusicGen large | musicgen | 3.3 | false | null | null | null | null | null | null | null | 2023-06 | null | null | audio | CC-BY-NC-4.0 | no | Model weights are CC-BY-NC 4.0 (non-commercial); the AudioCraft code is MIT. | {
"subtype": "music",
"backbone_params_b": 3.3,
"precision": "fp32",
"recommended": "{\"gb\":20,\"quant\":\"fp32\",\"source\":\"https:\\/\\/github.com\\/facebookresearch\\/audiocraft\\/blob\\/main\\/docs\\/MUSICGEN.md\",\"synthesis\":true}",
"cpu_ok": false,
"device_classes": [
"nvidia"
],
"tools": ... | [
"https://huggingface.co/facebook/musicgen-large",
"https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
] | Largest MusicGen (3.3B). The fp32 weights alone are ~13GB; a 24GB GPU is recommended, and peak consumed is ~20GB (synthesis from the weight size plus activations, no official figure). Weights are CC-BY-NC (non-commercial). Sources: HF card, AudioCraft MusicGen docs. |
stable-audio-open | Stable Audio Open 1.0 | stable-audio | 1.3 | false | null | null | null | null | null | null | null | 2024-06 | null | null | audio | Stability Community License | conditional | Free for commercial use under $1M annual revenue; enterprise license above. Cannot be used to train other generative models. | {
"subtype": "music",
"backbone_params_b": 1.06,
"precision": "fp32 / fp16",
"recommended": "{\"gb\":15,\"quant\":\"fp32\",\"source\":\"https:\\/\\/arxiv.org\\/html\\/2407.14358v1\"}",
"cpu_ok": false,
"device_classes": [
"nvidia",
"amd"
],
"tools": [
"stable-audio-tools",
"diffusers",
... | [
"https://huggingface.co/stabilityai/stable-audio-open-1.0",
"https://arxiv.org/html/2407.14358v1",
"https://huggingface.co/stabilityai/stable-audio-open-1.0/blob/main/LICENSE.md"
] | Latent-diffusion music model (1.06B DiT + autoencoder + T5-base, ~1.3B total) generating up to 47s of 44.1kHz stereo. The diffusion phase uses ~5.9GB VRAM; the decoder peaks at ~14.5GB (measured on an RTX 3090), so budget ~15GB unless you use chunked decoding. Stability Community License: free commercial use under $1M ... |
apple-m1-8gb | Apple M1 (8GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/111883",
"https://developer.apple.com/forums/thread/732035",
"https://blog.peddals.com/en/fine-tune-vram-size-of-mac-for-llm/",
"https://en.wikipedia.org/wiki/Apple_M1",
"https://www.apple.com/newsroom/2020/03/new-macbook-air-has-more-to-love-and-is-now-just-999/",
"https:... | Base MacBook Air M1 config. recommendedMaxWorkingSetSize ~66% of RAM for <64GB configs = ~5.3–5.5GB usable for model weights. macOS needs ~2–3GB for OS overhead on top. Practical limit: 3–4B parameter models at Q4_K_M. |
apple-m2-16gb | Apple M2 (16GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/111869",
"https://developer.apple.com/forums/thread/732035",
"https://stencel.io/posts/apple-silicon-limitations-with-usage-on-local-llm%20.html",
"https://www.apple.com/newsroom/2022/06/apple-unveils-m2-with-breakthrough-performance-and-capabilities/",
"https://www.apple.co... | Available in MacBook Pro 13-inch M2 (2022) and MacBook Air M2. recommendedMaxWorkingSetSize ~66% for <64GB = ~10.5GB. Fits 7–8B models at Q4_K_M comfortably. |
apple-m3-18gb | Apple M3 Pro (18GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/117736",
"https://developer.apple.com/forums/thread/732035",
"https://blog.peddals.com/en/fine-tune-vram-size-of-mac-for-llm/",
"https://support.apple.com/en-us/118551",
"https://www.apple.com/newsroom/2024/03/apple-unveils-the-new-13-and-15-inch-macbook-air-with-the-powerfu... | Base M3 Pro config in MacBook Pro 14/16-inch 2023. recommendedMaxWorkingSetSize ~66% for <64GB = ~12GB. Fits 13B models at Q4_K_M, 7B with headroom. |
apple-m4-16gb | Apple M4 (16GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121555",
"https://support.apple.com/en-us/121552",
"https://developer.apple.com/forums/thread/732035",
"https://support.apple.com/en-us/122209",
"https://www.apple.com/newsroom/2025/03/apple-introduces-the-new-macbook-air-with-the-m4-chip-and-a-sky-blue-color/",
"https://s... | Base Mac mini M4 (2024) and MacBook Pro 14-inch M4 config. recommendedMaxWorkingSetSize ~66% for <64GB = ~10.5GB. MLX backend now available in Ollama. Fits 7–8B models at Q4_K_M. |
apple-m4-24gb | Apple M4 (24GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121555",
"https://developer.apple.com/forums/thread/732035",
"https://stencel.io/posts/apple-silicon-limitations-with-usage-on-local-llm%20.html",
"https://support.apple.com/en-us/122209",
"https://www.apple.com/newsroom/2025/03/apple-introduces-the-new-macbook-air-with-the-... | Configurable Mac mini M4 (2024). recommendedMaxWorkingSetSize ~66% for <64GB = ~16GB. Fits 13B models at Q4_K_M comfortably. |
apple-m4-pro-24gb | Apple M4 Pro (24GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121553",
"https://support.apple.com/en-us/121555",
"https://developer.apple.com/forums/thread/732035",
"https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/",
"https://www.apple.com/newsroom/2024/10/new-macbook-pro-features-m4-family-of-chips-and-apple-... | Base M4 Pro config in Mac mini and MacBook Pro 14/16-inch 2024. Same 66% rule applies for <64GB = ~16GB. Higher memory bandwidth than base M4 chip benefits throughput. |
apple-m4-pro-48gb | Apple M4 Pro (48GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121553",
"https://support.apple.com/en-us/121555",
"https://blog.peddals.com/en/fine-tune-vram-size-of-mac-for-llm/",
"https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/",
"https://www.apple.com/newsroom/2024/10/new-macbook-pro-features-m4-family-of-c... | Top M4 Pro config in MacBook Pro 14/16-inch 2024 and Mac mini. recommendedMaxWorkingSetSize ~66% for <64GB = ~32GB. Fits 34B models at Q4_K_M. Strong sweet-spot for local LLM. |
apple-m4-max-64gb | Apple M4 Max (64GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121553",
"https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/",
"https://developer.apple.com/forums/thread/732035",
"https://support.apple.com/en-us/102027"
] | Base M4 Max config. At 64GB, rule shifts to ~75% (per Apple Metal docs) = ~48GB usable. 546GB/s memory bandwidth. Comfortably fits 34B models; 70B at lower quant. |
apple-m4-max-128gb | Apple M4 Max (128GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/121553",
"https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/",
"https://stencel.io/posts/apple-silicon-limitations-with-usage-on-local-llm%20.html",
"https://support.apple.com/en-us/102027"
] | Top M4 Max config. 75% rule at ≥64GB = 96GB confirmed usable (matches 128GB M1 Ultra precedent from stencel.io). Fits 70B models at full precision or Q8; handles 405B at lower quants. |
apple-m3-ultra-256gb | Apple M3 Ultra (256GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/122211",
"https://www.apple.com/newsroom/2025/03/apple-reveals-m3-ultra-taking-apple-silicon-to-a-new-extreme/",
"https://stencel.io/posts/apple-silicon-limitations-with-usage-on-local-llm%20.html",
"https://www.apple.com/mac-studio/specs/",
"https://support.apple.com/en-us/... | Max config Mac Studio M3 Ultra (2025). 75% rule at ≥64GB = 192GB usable. The 512GB config was discontinued in March 2026; 256GB remains available. Supports the largest open-weight models (405B+) at reasonable quants. |
nvidia-rtx-3060-12gb | Nvidia GeForce RTX 3060 (12GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3060-3060ti/",
"https://marketplace.nvidia.com/en-us/consumer/graphics-cards/msi-gaming-geforce-rtx-3060-12gb-15-gbps-gdrr6-192-bit-hdmi-dp-pcie-4-torx-twin-fan-ampere-oc-graphics-card/",
"https://en.wikipedia.org/wiki/GeForce_RTX_30_series",
... | 12GB GDDR6 on 192-bit bus. Unusually high VRAM for its tier. ~1GB reserved for driver/OS = ~11GB usable. Fits 7B Q4_K_M (4.1GB) and 13B Q4_K_M (7.4GB) with room. CUDA support via Ampere architecture. |
nvidia-rtx-4060-ti-16gb | Nvidia GeForce RTX 4060 Ti (16GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4060-4060ti/",
"https://www.techspot.com/specs/gpu/280961-nvidia-geforce-rtx-4060-ti-16gb.html",
"https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
"https://videocardz.com/newz/nvidia-geforce-rtx-4060-ti-16gb-to-feature-ad106-351-gpu-and-1... | 16GB GDDR6 on 128-bit bus (narrow bandwidth: 288 GB/s). ~1GB reserved = ~15GB usable. Bandwidth bottleneck limits throughput vs wider-bus alternatives. Fits 13B Q4_K_M easily; 34B at low quant. |
nvidia-rtx-4070-12gb | Nvidia GeForce RTX 4070 (12GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4070/",
"https://www.nvidia.com/en-us/geforce/news/rtx-40-series-vram-video-memory-explained/",
"https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
"https://videocardz.com/newz/nvidia-geforce-rtx-4070-specs-and-599-pricing-confirmed-186w-av... | 12GB GDDR6X on 192-bit bus (~504 GB/s). ~1GB reserved = ~11GB usable. Better bandwidth than 4060 Ti 16GB despite less VRAM. Fits 7B comfortably, 13B Q4 squeezed. |
nvidia-rtx-4080-16gb | Nvidia GeForce RTX 4080 (16GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4080/",
"https://www.nvidia.com/en-us/geforce/news/rtx-40-series-vram-video-memory-explained/",
"https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
"https://www.rxelectronics.nz/datasheet/a4/geforce-rtx-4080.pdf",
"https://www.pny.com/fil... | 16GB GDDR6X on 256-bit bus (~716 GB/s). ~1GB reserved = ~15GB usable. Good bandwidth for 13B models; 34B models require offloading. Strong for inference throughput. |
nvidia-rtx-4090-24gb | Nvidia GeForce RTX 4090 (24GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/",
"https://www.nvidia.com/en-us/geforce/news/rtx-40-series-vram-video-memory-explained/",
"https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
"https://wccftech.com/roundup/nvidia-geforce-rtx-4090/"
] | 24GB GDDR6X on 384-bit bus (~1008 GB/s). ~1GB reserved = ~23GB usable. Flagship Ada Lovelace. Fits 34B Q4_K_M (19GB); 70B requires offloading or very low quant. Best single-GPU consumer option for throughput. |
nvidia-rtx-5090-32gb | Nvidia GeForce RTX 5090 (32GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/",
"https://www.spheron.network/blog/nvidia-rtx-5090-specs/",
"https://en.wikipedia.org/wiki/GeForce_RTX_50_series"
] | 32GB GDDR7 on 512-bit bus (~1792 GB/s). ~1GB reserved = ~31GB usable. Blackwell architecture. Fits 34B Q8 and pushes toward 70B Q4_K_M (~35–38GB, tight). Best single-GPU consumer option as of 2026. |
nvidia-rtx-3090-24gb | Nvidia GeForce RTX 3090 (24GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090/",
"https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/",
"https://en.wikipedia.org/wiki/GeForce_RTX_30_series",
"https://wccftech.com/nvidia-geforce-rtx-3090-24-gb-official-launch-price-specs-performance/"
] | 24GB GDDR6X on 384-bit bus (~936 GB/s). ~1GB reserved = ~23GB usable. Ampere architecture. Excellent value-per-GB for LLM inference; same VRAM ceiling as RTX 4090 at lower bandwidth. Fits 34B Q4_K_M. |
amd-rx-7900-xtx-24gb | AMD Radeon RX 7900 XTX (24GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.notebookcheck.net/AMD-Radeon-RX-7900-XTX-with-24-GB-VRAM-review-Already-available-for-less-than-1000-Euros.810630.0.html",
"https://www.asus.com/us/motherboards-components/graphics-cards/asus/rx7900xtx-24g/techspec/",
"https://bestgpuforllm.com/articles/nvidia-vs-amd-for-llm/",
"https://www.amd.c... | 24GB GDDR6 on 384-bit bus (~960 GB/s). ~1GB reserved = ~23GB usable. RDNA 3 architecture. ROCm support on Linux; ROCm on Windows is experimental/limited. LLM inference quality behind CUDA in software maturity. Fits 34B Q4_K_M. Windows users typically prefer llama.cpp with ROCm or DirectML backend. |
amd-ryzen-ai-halo-128gb | AMD Ryzen AI Halo (128GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.amd.com/en/blogs/2026/amd-powers-next-generation-agent-computers-with-new-ryzen-ai-hal.html",
"https://www.amd.com/en/blogs/2025/faqs-amd-variable-graphics-memory-vram-ai-model-sizes-quantization-mcp-more.html",
"https://frame.work/blog/framework-desktop-deep-dive-ryzen-ai-max",
"https://www.cnx-... | AMD's Ryzen AI Halo Developer Platform (Strix Halo, Ryzen AI Max+ 395): 128GB LPDDR5X-8000 on a 256-bit bus, 256 GB/s. Unified memory: up to 96GB convertible to VRAM via AMD Variable Graphics Memory, per AMD's FAQ. Chip cTDP 45-120W; the dev box runs 120W per CNX's spec table. $3,999.99 at Micro Center, in stores 2026-... |
laptop-8gb | 8GB RAM Laptop (CPU/iGPU only) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://ollama.com/library/llama3.2:1b",
"https://ollama.com/library/phi3:mini",
"https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF",
"https://en.wikipedia.org/wiki/DDR5_SDRAM",
"https://www.notebookcheck.net/Intel-Core-Ultra-5-125U-Processor-Benchmarks-and-Specs.783351.0.html",
"https://www.am... | OS overhead ~2-3GB (Windows/macOS), leaving ~5GB usable. Practical ceiling is 1B-3B Q4_K_M models. Llama 3.2 1B Q4_K_M ~0.8GB file, needs ~1.2GB at runtime. Phi-3 Mini 3.8B Q4_K_M ~2.4GB file, needs ~3.6GB at runtime, fits. 7B+ models will OOM. Usable figure is community estimate, not vendor spec. |
laptop-16gb | 16GB RAM Laptop (CPU/iGPU only) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://ollama.com/library/llama3.1:8b",
"https://ollama.com/library/mistral:7b",
"https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
"https://en.wikipedia.org/wiki/DDR5_SDRAM",
"https://www.notebookcheck.net/Intel-Core-Ultra-7-155H-Processor-Benchmarks-and-Specs.783323.0.html",
"https://... | OS overhead ~2-4GB, leaving ~12GB usable. Practical ceiling is 7B-8B Q4_K_M models. Llama 3.1 8B Q4_K_M ~4.7GB file, needs ~7GB at runtime. Mistral 7B Q4_K_M ~4.1GB file, needs ~6GB at runtime. 13B models are marginal and will be slow. Usable figure is community estimate, not vendor spec. |
laptop-32gb | 32GB RAM Laptop (CPU/iGPU only) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://ollama.com/library/llama3.1:70b",
"https://ollama.com/library/mixtral:8x7b",
"https://huggingface.co/bartowski/Meta-Llama-3.1-70B-Instruct-GGUF",
"https://en.wikipedia.org/wiki/DDR5_SDRAM",
"https://www.notebookcheck.net/Intel-Core-Ultra-7-155H-Processor-Benchmarks-and-Specs.783323.0.html",
"http... | OS overhead ~2-4GB, leaving ~28GB usable. Practical ceiling is 13B-34B Q4_K_M models comfortably; 70B at Q2_K (lossy) marginally fits (~28GB). Llama 3.1 13B Q4_K_M ~7.8GB file, needs ~11.7GB. Llama 3.3 70B Q4_K_M ~43GB, does NOT fit; Q2_K ~24GB fits but quality degraded. Usable figure is community estimate, not vendor ... |
iphone-15-pro | iPhone 15 Pro | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/apple_iphone_15_pro-12178.php",
"https://wccftech.com/iphone-15-pro-ram-type-lpddr5-confirmed/",
"https://www.techinsights.com/blog/apple-iphone-15-pro-teardown",
"https://www.cpu-monkey.com/en/cpu-apple_a17_pro",
"https://support.apple.com/en-us/111829",
"https://forums.macrumor... | 8GB LPDDR5 unified memory confirmed via TechInsights teardown (Micron D1b LPDDR5 chips) and GSMArena. iOS reserves ~3-4GB for OS+system, leaving ~4-4.5GB usable for model weights (community estimate). Practical ceiling 1B-3B Q4_K_M on-device; Apple Intelligence models (Llama 3B class) run at ~3-4GB. CoreML path most me... |
iphone-16 | iPhone 16 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/apple_iphone_16-12557.php",
"https://www.macrumors.com/2024/09/09/all-iphone-16-models-equipped-with-8gb-of-ram/",
"https://www.apple.com/iphone-16/specs/",
"https://en.wikipedia.org/wiki/Apple_A18",
"https://en.wikipedia.org/wiki/IPhone_16"
] | All iPhone 16 models confirmed 8GB RAM per MacRumors (Sep 2024, citing Apple Intelligence requirement). Apple specs page confirms 8GB. Unified memory architecture; iOS reserves ~3-4GB, leaving ~4-4.5GB usable (community estimate). Identical LLM capability envelope to iPhone 15 Pro. On-device model ceiling: 1B-3B Q4_K_M... |
iphone-16-pro | iPhone 16 Pro | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/apple_iphone_16_pro-12560.php",
"https://www.macrumors.com/2024/09/09/all-iphone-16-models-equipped-with-8gb-of-ram/",
"https://www.apple.com/iphone-16-pro/specs/",
"https://en.wikipedia.org/wiki/Apple_A18",
"https://versus.com/en/apple-a18-vs-apple-a18-pro",
"https://www.apple.c... | 8GB unified memory confirmed (same MacRumors source as iPhone 16, all 4 iPhone 16 models have 8GB). A18 Pro chip with 6-core GPU gives higher inference throughput vs A16/A17 but same memory envelope. iOS overhead ~3-4GB; usable ~4.5GB (community estimate). Practical ceiling 1B-4B Q4_K_M. Off Grid app (llama.cpp+Metal) ... |
ipad-pro-m4-16gb | iPad Pro M4 (16GB, 1TB/2TB config) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/119891",
"https://www.apple.com/ipad-pro/specs/",
"https://www.gsmarena.com/apple_ipad_pro_13_(2024)-12342.php",
"https://support.apple.com/en-us/119892",
"https://www.apple.com/newsroom/2024/05/apple-unveils-stunning-new-ipad-pro-with-m4-chip-and-apple-pencil-pro/",
"http... | iPad, not iPhone, categorized as 'iphone' due to schema enum limitation. 16GB unified memory only available on 1TB and 2TB storage configurations per Apple Support KB article (support.apple.com/en-us/119891). 256GB/512GB configs ship with 8GB. M4 chip with 10-core GPU. iPadOS overhead ~3-4GB; ~12GB usable for model wei... |
pixel-9-pro | Google Pixel 9 Pro | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/google_pixel_9_pro-12424.php",
"https://www.androidpolice.com/google-pixel-9-pro-ram-partition-gemini/",
"https://mlc.ai/blog/2024/05/08/mlc-llm-android",
"https://comparigon.com/mobilesocs/tensor-g4",
"https://9to5google.com/2024/08/13/pixel-9-pro-and-pro-xl/",
"https://www.gsma... | 16GB LPDDR5X confirmed via GSMArena. Android Police investigation confirmed 2.6GB hardware-partitioned for Google Gemini/on-device ML features, not available to apps. Android OS overhead ~3-4GB additional; usable estimate ~10-10.5GB (community estimate). Practical ceiling: 7B Q4_K_M fits (~7GB runtime); 13B Q4_K_M tigh... |
samsung-s24-ultra | Samsung Galaxy S24 Ultra | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/samsung_galaxy_s24_ultra-12419.php",
"https://www.gsmarena.com/samsung_galaxy_s24_ultra-review-2754p5.php",
"https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
"https://comparigon.com/mobilesocs/snapdragon-8-gen-3",
"https://www.gsmarena.com/samsung_galaxy_s24_ultra... | 12GB LPDDR5X at 4800MHz confirmed via GSMArena review page (explicitly states RAM type and speed). All S24 Ultra configs are 12GB (no 16GB tier unlike S25 Ultra). Snapdragon 8 Gen 3 (US). Android OS overhead ~3-4GB; usable ~8-8.5GB (community estimate). Practical ceiling: 7B Q4_K_M (~7GB runtime) fits; 13B Q4_K_M (~11.... |
samsung-s25-ultra-16gb | Samsung Galaxy S25 Ultra (16GB, 1TB config only) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/samsung_galaxy_s25_ultra-12750.php",
"https://www.sammyfans.com/2025/01/22/galaxy-s25-ultra-ram-lpddr5x-micron/",
"https://www.sammyguru.com/samsung-galaxy-s25-ultra-ram-specs-confirmed/",
"https://gadgetversus.com/processor/qualcomm-sm8750-ac-snapdragon-8-elite-specs/",
"https://w... | 16GB LPDDR5X (Micron, 12nm process) confirmed via SammyFans and SammyGuru. CRITICAL: 16GB RAM is ONLY available on the 1TB storage configuration. 256GB and 512GB configs ship with 12GB RAM. Android OS overhead ~3-4GB; usable ~12GB (community estimate). Snapdragon 8 Elite. Practical ceiling: 7B-13B Q4_K_M. Llama 3.1 8B ... |
android-generic-8gb | Generic Android Phone (8GB RAM) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://ollama.com/library/llama3.2:1b",
"https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF",
"https://github.com/a-ghorbani/pocketpal-ai"
] | Representative of mid-range Android devices (e.g., Pixel 8a, OnePlus 12R, Samsung A55). RAM type varies (LPDDR4X on lower-end, LPDDR5 on higher-end). Android OS overhead ~3-4GB; usable ~4-4.5GB (community estimate). Practical ceiling: 1B-3B Q4_K_M only. Llama 3.2 1B Q4_K_M (~1.2GB runtime) fits well. Phi-3 Mini 3.8B Q4... |
android-generic-12gb | Generic Android Phone (12GB RAM) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://ollama.com/library/llama3.1:8b",
"https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
"https://github.com/a-ghorbani/pocketpal-ai"
] | Representative of upper-mid and flagship Android devices (e.g., OnePlus 12, Pixel 9, Samsung S24+). LPDDR5 or LPDDR5X depending on SoC. Android OS overhead ~3-4GB; usable ~8-8.5GB (community estimate). Practical ceiling: 7B Q4_K_M models. Llama 3.1 8B Q4_K_M (~7GB runtime) fits with ~1.5GB headroom. 13B Q4_K_M (~11.7GB... |
apple-m5-16gb | Apple M5 (16GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/125405",
"https://www.apple.com/macbook-pro/specs/",
"https://developer.apple.com/forums/thread/732035",
"https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/",
"https://www.apple.com/newsroom/2026/03/apple-introduc... | Base MacBook Pro 14-inch M5 (2025). recommendedMaxWorkingSetSize ~66% for <64GB = ~10.5GB usable. Fits 7-8B models at Q4_K_M. M5 adds a Neural Accelerator per GPU core for faster prompt processing. |
apple-m5-32gb | Apple M5 (32GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/125405",
"https://www.apple.com/macbook-pro/specs/",
"https://developer.apple.com/forums/thread/732035",
"https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/",
"https://www.apple.com/newsroom/2026/03/apple-introduc... | Top base-M5 config (MacBook Pro 14-inch M5, 2025); 32GB is the max on the base M5 chip. ~66% for <64GB = ~21GB usable. Fits 14B at Q4_K_M comfortably, 32B tight. |
apple-m5-pro-48gb | Apple M5 Pro (48GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/126318",
"https://www.apple.com/macbook-pro/specs/",
"https://blog.peddals.com/en/fine-tune-vram-size-of-mac-for-llm/",
"https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/"
] | M5 Pro (MacBook Pro 14/16-inch). Configurable to 64GB; 48GB config shown. ~66% for <64GB = ~32GB usable. ~307GB/s memory bandwidth. Fits 34B at Q4_K_M. |
apple-m5-max-128gb | Apple M5 Max (128GB) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://support.apple.com/en-us/126318",
"https://support.apple.com/en-us/126319",
"https://www.apple.com/macbook-pro/specs/",
"https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/"
] | M5 Max top config. 75% rule at >=64GB = ~96GB usable. ~614GB/s memory bandwidth. Fits 70B at Q8, or 235B-class MoE at Q4 with room to spare. |
iphone-17 | iPhone 17 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.macrumors.com/2025/09/09/iphone-17-pro-iphone-air-ram-amounts/",
"https://www.apple.com/iphone-17/specs/",
"https://en.wikipedia.org/wiki/Apple_A19",
"https://en.wikipedia.org/wiki/IPhone_17"
] | iPhone 17 confirmed 8GB RAM (MacRumors, Sep 2025). A19 chip. iOS reserves ~3-4GB; ~4.5GB usable for weights (community estimate). On-device ceiling 1B-4B Q4_K_M. |
iphone-17-pro | iPhone 17 Pro | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.macrumors.com/2025/09/09/iphone-17-pro-iphone-air-ram-amounts/",
"https://www.apple.com/iphone-17-pro/specs/",
"https://en.wikipedia.org/wiki/Apple_A19",
"https://wccftech.com/a19-pro-geekbench-6-performance-per-watt-analysis/"
] | iPhone 17 Pro confirmed 12GB RAM (MacRumors, Sep 2025), up from 8GB on iPhone 16 Pro. A19 Pro. iOS overhead ~3.5-4GB; ~8GB usable (community estimate). The extra RAM lifts the on-device ceiling toward 7-8B Q4_K_M, a first for iPhone. |
iphone-air | iPhone Air | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.macrumors.com/2025/09/09/iphone-17-pro-iphone-air-ram-amounts/",
"https://www.apple.com/iphone-air/specs/",
"https://en.wikipedia.org/wiki/IPhone_Air",
"https://www.gsmarena.com/apple_iphone_17_air-13502.php"
] | iPhone Air confirmed 12GB RAM (MacRumors, Sep 2025). A19 Pro. iOS overhead ~3.5-4GB; ~8GB usable (community estimate). On-device ceiling ~7-8B Q4_K_M. |
pixel-10-pro | Google Pixel 10 Pro | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/google_pixel_10_pro_5g-13987.php",
"https://store.google.com/product/pixel_10_pro_specs",
"https://gadgetversus.com/processor/google-tensor-g5-specs/",
"https://store.google.com/product/pixel_10_pro?hl=en-US",
"https://9to5google.com/2025/08/20/pixel-10-tensor-g5/"
] | 16GB LPDDR5X confirmed via GSMArena. Tensor G5. Memory is reserved for on-device Gemini Nano plus ~3-4GB Android OS overhead; usable ~10-10.5GB (community estimate). Practical ceiling 7B Q4_K_M; 13B tight. |
samsung-s26-ultra | Samsung Galaxy S26 Ultra (16GB, 1TB config) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://www.gsmarena.com/samsung_galaxy_s26_ultra_5g-14320.php",
"https://www.sammyfans.com/2026/03/13/samsung-galaxy-s26-ultra-details/",
"https://www.notebookcheck.net/Qualcomm-Snapdragon-8-Elite-Gen-5-Processor-Benchmarks-and-Specs.1123169.0.html",
"https://www.sammobile.com/news/samsung-galaxy-s26-ultra-... | 16GB LPDDR5X on the 1TB config (256/512GB ship 12GB) per GSMArena and SammyFans. Snapdragon 8 Elite Gen 5. Android overhead ~3-4GB; usable ~12GB (community estimate). Fits 7B-13B Q4_K_M. MLC-LLM + Adreno for best throughput. |
stable-diffusion-1-5 | Stable Diffusion 1.5 | stable-diffusion | 0.86 | false | null | null | null | null | null | null | null | 2022-10 | null | null | image | CreativeML OpenRAIL-M | yes | Use-based restrictions (no illegal or harmful content); no revenue cap. | null | [
"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5",
"https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations",
"https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors"
] | UNet 0.86B (~0.98B with the CLIP-L text encoder; no T5). The single CLIP-L encoder is ~0.25GB so all components stay resident and peak VRAM is the sum: UNet ~1.7GB + CLIP-L ~0.25GB + VAE ~0.17GB plus activations is ~3.7GB at 512x512, the widely-cited 4GB minimum. Runs on 2GB with AUTOMATIC1111 --lowvram, slowly. Source... |
sdxl-1-0 | Stable Diffusion XL 1.0 | stable-diffusion | 2.6 | false | null | null | null | null | null | null | null | 2023-07 | null | null | image | CreativeML OpenRAIL++-M | yes | Use-based restrictions; no revenue cap. | null | [
"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0",
"https://stability.ai/news-updates/sdxl-09-stable-diffusion",
"https://github.com/Comfy-Org/ComfyUI/issues/2855",
"https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimum-SDXL-Usage"
] | UNet 2.6B (3.5B with CLIP-L + OpenCLIP-G; no T5). The two CLIP encoders total ~1.6GB and stay resident, so peak VRAM is the sum: ~7.5GB measured at 1024x1024 on an 8GB card, matching Stability's stated 8GB minimum. Runs on 4GB with AUTOMATIC1111 --lowvram, slowly. An optional refiner adds a second ~6GB UNet. Sources: S... |
stable-diffusion-3-5-large | Stable Diffusion 3.5 Large | stable-diffusion | 8.1 | false | null | null | null | null | null | null | null | 2024-10 | null | null | image | Stability Community License | conditional | Free for commercial use under $1M annual revenue; an enterprise license is required above that. | null | [
"https://stability.ai/news-updates/introducing-stable-diffusion-3-5",
"https://huggingface.co/city96/stable-diffusion-3.5-large-gguf",
"https://stability.ai/news-updates/stable-diffusion-35-models-optimized-with-tensorrt-deliver-2x-faster-performance-and-40-less-memory-on-nvidia-rtx-gpus",
"https://huggingfac... | 8.1B MMDiT with three text encoders (CLIP-L, OpenCLIP-G, T5-XXL). The 9.8GB T5-XXL is offloaded to CPU after prompt-encoding, so peak VRAM tracks the backbone, not the sum. At Q4 GGUF (4.77GB backbone) with T5 offloaded, peak is ~7GB (synthesis from city96 component sizes plus the diffusers offload behavior, not a sing... |
flux-1-dev | FLUX.1 dev | flux | 12 | false | null | null | null | null | null | null | null | 2024-08 | null | null | image | FLUX.1-dev Non-Commercial License | no | Weights are non-commercial. Generated images may be used commercially, but using the model inside a product needs a separate Black Forest Labs license. | null | [
"https://huggingface.co/black-forest-labs/FLUX.1-dev",
"https://huggingface.co/city96/FLUX.1-dev-gguf",
"https://huggingface.co/city96/FLUX.1-dev-gguf/discussions/9",
"https://huggingface.co/docs/diffusers/optimization/memory",
"https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md"
] | 12B DiT with CLIP-L and a 9.8GB T5-XXL text encoder. T5 is offloaded to CPU after prompt-encoding, so the Q4 GGUF backbone (6.8GB) drives a denoising peak of ~6.5GB, measured at 6.4GB on an RTX 2080 8GB in Forge. All-components-resident at bf16 is ~33GB. Sequential CPU offload runs on ~3GB, very slowly. Non-commercial ... |
flux-1-schnell | FLUX.1 schnell | flux | 12 | false | null | null | null | null | null | null | null | 2024-08 | null | null | image | Apache-2.0 | yes | Apache-2.0; the only FLUX.1 variant licensed for commercial products. | null | [
"https://huggingface.co/black-forest-labs/FLUX.1-schnell",
"https://huggingface.co/city96/FLUX.1-schnell-gguf",
"https://huggingface.co/docs/diffusers/optimization/memory"
] | The same 12B DiT as FLUX.1 dev, distilled to 1-4 step generation and licensed Apache-2.0 (the only commercially-usable FLUX.1 variant). T5-XXL is offloaded after prompt-encoding, so the Q4 GGUF backbone peaks at ~6.5GB during denoising. All-resident bf16 is ~33GB. Sources: BFL schnell model card, city96 FLUX.1-schnell ... |
qwen-image | Qwen-Image | qwen | 20 | false | null | null | null | null | null | null | null | 2025-08 | null | null | image | Apache-2.0 | yes | null | null | [
"https://huggingface.co/Qwen/Qwen-Image",
"https://huggingface.co/city96/Qwen-Image-gguf",
"https://docs.comfy.org/tutorials/image/qwen/qwen-image",
"https://sandner.art/qwen-image-and-edit-local-gguf-generations-with-lightning/",
"https://github.com/QwenLM/Qwen-Image"
] | 20B MMDiT with a 7B Qwen2.5-VL text encoder. The encoder is offloaded after prompt-encoding, so the Q4_K_M GGUF backbone (13.1GB) sustains ~14GB during denoising, validated on a 16GB RTX A4000. Full bf16 with the encoder resident is ~57GB (multi-GPU). Nunchaku SVDQuant 4-bit can offload to ~3GB VRAM, slowly. Apache-2.0... |
llama-3.1-8b | Llama 3.1 8B | Llama | 8 | false | null | null | null | null | null | null | null | 2024-07 | null | null | null | Llama 3.1 Community | conditional | Llama 3.1 Community License: free under 700M MAU. | null | [
"https://ollama.com/library/llama3.1",
"https://ollama.com/library/llama3.1/tags",
"https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Q4_K_M and Q8_0 sizes from bartowski HF repo. Default ollama tag pulls 4.9GB Q4_K_M. Min RAM estimate includes model + small KV cache overhead. |
llama-3.3-70b | Llama 3.3 70B | Llama | 70 | false | null | null | null | null | null | null | null | 2024-12 | null | null | null | Llama 3.3 Community | conditional | Llama 3.3 Community License: free under 700M MAU. | null | [
"https://ollama.com/library/llama3.3",
"https://ollama.com/library/llama3.3/tags",
"https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released December 6, 2024. Delivers near-405B performance at 70B cost. Q4_K_M and Q8_0 sizes from bartowski HF repo cross-validated against Ollama tags page (43GB and 75GB displayed). |
llama-3.2-3b | Llama 3.2 3B | Llama | 3 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Llama 3.2 Community | conditional | Llama 3.2 Community License: free under 700M MAU. | null | [
"https://ollama.com/library/llama3.2",
"https://ollama.com/library/llama3.2/tags",
"https://huggingface.co/unsloth/Llama-3.2-3B-Instruct-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released September 25, 2024. Text-only edge model. Ollama default tag is 3B at 2.0GB. HF unsloth repo confirms 2.02GB Q4_K_M and 3.42GB Q8_0. |
llama-3.2-1b | Llama 3.2 1B | Llama | 1 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Llama 3.2 Community | conditional | Llama 3.2 Community License: free under 700M MAU. | null | [
"https://ollama.com/library/llama3.2",
"https://ollama.com/library/llama3.2/tags",
"https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released September 25, 2024. Smallest Llama model for edge/on-device use. Ollama tag shows 1.3GB (Q8_0 equivalent default). HF unsloth confirms Q4_K_M at 808MB (~0.81GB) and Q8_0 at 1.32GB. |
mistral-7b | Mistral 7B | Mistral | 7 | false | null | null | null | null | null | null | null | 2024-05 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/mistral",
"https://ollama.com/library/mistral/tags",
"https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF",
"https://lmarena.ai/leaderboard"
] | Latest is v0.3 (released May 2024), adds function calling and extended vocabulary. Ollama default tag shows 4.4GB; bartowski HF repo gives precise 4.37GB Q4_K_M and 7.70GB Q8_0 for v0.3. |
mistral-small-3-24b | Mistral Small 3 24B | Mistral | 24 | false | null | null | null | null | null | null | null | 2025-06 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/mistral-small3.2",
"https://ollama.com/library/mistral-small3.2/tags",
"https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF",
"https://docs.mistral.ai/models/model-cards/mistral-small-3-2-25-06",
"https://lmarena.ai/leaderboard"
] | Mistral Small 3.2 released June 20, 2025. Latest in the Mistral Small 3 line. Context 128K confirmed by official Mistral docs. Q4_K_M=14.33GB and Q8_0=25.05GB from bartowski HF repo (mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF). Ollama page shows 15GB for the default tag. |
phi-4-14b | Phi-4 14B | phi | 14 | false | null | null | null | null | null | null | null | 2024-12 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/phi4",
"https://ollama.com/library/phi4/tags",
"https://huggingface.co/bartowski/phi-4-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released December 12, 2024 by Microsoft. Strong math and reasoning. Default context 16K. Ollama shows 9.1GB; bartowski HF gives precise 9.05GB Q4_K_M and 15.58GB Q8_0. |
phi-4-mini-3.8b | Phi-4-mini 3.8B | phi | 3.8 | false | null | null | null | null | null | null | null | 2025-02 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/phi4-mini",
"https://ollama.com/library/phi4-mini/tags",
"https://huggingface.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF"
] | Released February 2025 by Microsoft. Dense decoder-only transformer with 200K vocabulary. 128K context. Q4_K_M=2.49GB and Q8_0=4.08GB from bartowski HF repo (microsoft_Phi-4-mini-instruct-GGUF). Requires Ollama 0.5.13+. |
gemma-2-9b | Gemma 2 9B | Gemma | 9 | false | null | null | null | null | null | null | null | 2024-06 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma2",
"https://ollama.com/library/gemma2/tags",
"https://huggingface.co/bartowski/gemma-2-9b-it-GGUF",
"https://lmarena.ai/leaderboard"
] | Released June 27, 2024. Uses sliding window attention with 8K context. Ollama default tag shows 5.4GB; bartowski HF repo gives precise 5.76GB Q4_K_M and 9.83GB Q8_0. |
gemma-2-27b | Gemma 2 27B | Gemma | 27 | false | null | null | null | null | null | null | null | 2024-06 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma2",
"https://ollama.com/library/gemma2/tags",
"https://huggingface.co/bartowski/gemma-2-27b-it-GGUF",
"https://lmarena.ai/leaderboard"
] | Released June 27, 2024. Uses sliding window attention with 8K context. Ollama default tag shows 16GB; bartowski HF repo gives precise 16.65GB Q4_K_M and 28.94GB Q8_0. |
gemma-3-4b | Gemma 3 4B | Gemma | 4 | false | null | null | null | null | null | null | null | 2025-03 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma3",
"https://ollama.com/library/gemma3/tags",
"https://huggingface.co/bartowski/google_gemma-3-4b-it-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released March 12, 2025. Multimodal (text + image), 140+ languages. Ollama shows 3.3GB default; bartowski HF (google_gemma-3-4b-it-GGUF) gives precise 2.49GB Q4_K_M and 4.13GB Q8_0. |
gemma-3-12b | Gemma 3 12B | Gemma | 12 | false | null | null | null | null | null | null | null | 2025-03 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma3",
"https://ollama.com/library/gemma3/tags",
"https://huggingface.co/bartowski/google_gemma-3-12b-it-GGUF",
"https://lmarena.ai/leaderboard",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released March 12, 2025. Multimodal (text + image), 140+ languages. Ollama shows 8.1GB default; bartowski HF (google_gemma-3-12b-it-GGUF) gives precise 7.30GB Q4_K_M and 12.51GB Q8_0. |
gemma-3-27b | Gemma 3 27B | Gemma | 27 | false | null | null | null | null | null | null | null | 2025-03 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma3",
"https://ollama.com/library/gemma3/tags",
"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF",
"https://lmarena.ai/leaderboard",
"https://aider.chat/docs/leaderboards/",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Released March 12, 2025. Multimodal (text + image), 140+ languages. Ollama shows 17GB default; bartowski HF (google_gemma-3-27b-it-GGUF) gives precise 16.55GB Q4_K_M and 28.71GB Q8_0. |
qwen2.5-7b | Qwen2.5 7B | Qwen2.5 | 7 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen2.5",
"https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Ollama shows 4.7GB (rounded); bartowski HF repo gives 4.68GB Q4_K_M and 8.10GB Q8_0. Default context 128K per official Qwen blog. Min RAM estimate adds ~1-2GB overhead over Q4_K_M size. |
qwen2.5-14b | Qwen2.5 14B | Qwen2.5 | 14 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen2.5",
"https://huggingface.co/bartowski/Qwen2.5-14B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Ollama shows 9.0GB (rounded); bartowski HF repo gives 8.99GB Q4_K_M and 15.70GB Q8_0. Default context 128K. |
qwen2.5-32b | Qwen2.5 32B | Qwen2.5 | 32 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen2.5",
"https://huggingface.co/bartowski/Qwen2.5-32B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Ollama shows 20GB (rounded); bartowski HF repo gives 19.85GB Q4_K_M and 34.82GB Q8_0. Default context 128K. |
qwen2.5-72b | Qwen2.5 72B | Qwen2.5 | 72 | false | null | null | null | null | null | null | null | 2024-09 | null | null | null | Qwen License | conditional | Qwen License: free for commercial use under 100M monthly active users. | null | [
"https://ollama.com/library/qwen2.5",
"https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/",
"https://lmarena.ai/leaderboard"
] | Ollama shows 47GB (rounded); bartowski HF repo gives 47.42GB Q4_K_M and 77.26GB Q8_0. Default context 128K. Requires multi-GPU or high-VRAM single GPU setup. |
qwen3-8b | Qwen3 8B | Qwen3 | 8 | false | null | null | null | null | null | null | null | 2025-04 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3/tags",
"https://huggingface.co/Qwen/Qwen3-8B-GGUF",
"https://qwenlm.github.io/blog/qwen3/",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Official Qwen/Qwen3-8B-GGUF HF repo lists Q4_K_M=5.03GB, Q8_0=8.71GB. Ollama library shows 5.2GB (rounded). Qwen3 released 2025-04-29. Context 32K native, extendable to 128K via YaRN. |
qwen3-14b | Qwen3 14B | Qwen3 | 14 | false | null | null | null | null | null | null | null | 2025-04 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3/tags",
"https://huggingface.co/Qwen/Qwen3-14B-GGUF",
"https://qwenlm.github.io/blog/qwen3/",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Official Qwen/Qwen3-14B-GGUF HF repo lists Q4_K_M=9GB, Q8_0=15.7GB. Confirmed by Ollama tags page (9.3GB rounded) and search cross-check. Context 32K native, 128K via YaRN. |
qwen3-32b | Qwen3 32B | Qwen3 | 32 | false | null | null | null | null | null | null | null | 2025-04 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3/tags",
"https://huggingface.co/Qwen/Qwen3-32B-GGUF",
"https://qwenlm.github.io/blog/qwen3/",
"https://lmarena.ai/leaderboard",
"https://aider.chat/docs/leaderboards/",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Official Qwen/Qwen3-32B-GGUF HF repo lists Q4_K_M=19.8GB, Q8_0=34.8GB. Cross-checked with Ollama tags page (20GB rounded). Context 32K native, 128K via YaRN. |
qwen3-30b-a3b | Qwen3 30B-A3B | Qwen3 | 30.5 | true | null | null | null | null | null | null | null | 2025-04 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3/tags",
"https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF",
"https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF",
"https://qwenlm.github.io/blog/qwen3/",
"https://lmarena.ai/leaderboard"
] | MoE: 30.5B total / 3.3B active (128 total experts, 8 activated per token). Q4_K_M=18.6GB and Q8_0=32.5GB from official unsloth/Qwen3-30B-A3B-GGUF HF repo, cross-confirmed with Qwen/Qwen3-30B-A3B-GGUF. Native context 32K, extendable to 128K via YaRN. Despite large Q4 size, inference is fast due to only 3.3B active param... |
deepseek-r1-distill-qwen-7b | DeepSeek-R1-Distill-Qwen 7B | DeepSeek-R1-Distill | 7 | false | null | null | null | null | null | null | null | 2025-01 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/deepseek-r1",
"https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF",
"https://github.com/deepseek-ai/DeepSeek-R1"
] | Distilled from DeepSeek-R1 onto Qwen2.5-7B base. Released 2025-01-20. Q4_K_M=4.68GB, Q8_0=8.10GB from bartowski HF repo. Ollama shows 4.7GB. Context 128K. |
deepseek-r1-distill-qwen-14b | DeepSeek-R1-Distill-Qwen 14B | DeepSeek-R1-Distill | 14 | false | null | null | null | null | null | null | null | 2025-01 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/deepseek-r1",
"https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF",
"https://github.com/deepseek-ai/DeepSeek-R1"
] | Distilled from DeepSeek-R1 onto Qwen2.5-14B base. Released 2025-01-20. Q4_K_M=8.99GB, Q8_0=15.70GB from bartowski HF repo. Ollama shows 9.0GB. Context 128K. |
deepseek-r1-distill-llama-8b | DeepSeek-R1-Distill-Llama 8B | DeepSeek-R1-Distill | 8 | false | null | null | null | null | null | null | null | 2025-01 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/deepseek-r1",
"https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF",
"https://github.com/deepseek-ai/DeepSeek-R1"
] | Distilled from DeepSeek-R1 onto Llama-3.1-8B base. Released 2025-01-20. Q4_K_M=4.92GB, Q8_0=8.54GB from bartowski HF repo. Ollama shows 5.2GB for the 8b tag. Context 128K. |
deepseek-r1-distill-qwen-32b | DeepSeek-R1-Distill-Qwen 32B | DeepSeek-R1-Distill | 32 | false | null | null | null | null | null | null | null | 2025-01 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/deepseek-r1",
"https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF",
"https://github.com/deepseek-ai/DeepSeek-R1"
] | Distilled from DeepSeek-R1 onto Qwen2.5-32B base. Released 2025-01-20. Q4_K_M=19.85GB, Q8_0=34.82GB from bartowski HF repo. Ollama shows 20GB. Context 128K. |
deepseek-v2-lite | DeepSeek-V2-Lite | DeepSeek-V2 | 16 | true | null | null | null | null | null | null | null | 2024-05 | null | null | null | DeepSeek License | yes | DeepSeek Model License permits commercial use. | null | [
"https://ollama.com/library/deepseek-v2",
"https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite",
"https://huggingface.co/mradermacher/DeepSeek-V2-Lite-GGUF",
"https://github.com/deepseek-ai/DeepSeek-V2"
] | MoE: 16B total / 2.4B active (2 shared + 64 routed experts per layer, 6 activated per token). Released 2024-05-16. Q4_K_M=10.4GB, Q8_0=16.8GB from mradermacher HF GGUF repo (bartowski does not host this model). Ollama deepseek-v2:16b=8.9GB uses different quantization. Context 32K native. Despite 16B total params, infer... |
smollm2-1.7b | SmolLM2 1.7B | SmolLM2 | 1.7 | false | null | null | null | null | null | null | null | 2024-11 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/smollm2",
"https://huggingface.co/bartowski/SmolLM2-1.7B-Instruct-GGUF",
"https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF",
"https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B",
"https://lmarena.ai/leaderboard"
] | Released by HuggingFace 2024-11-01. Q4_K_M=1.06GB, Q8_0=1.82GB from bartowski/SmolLM2-1.7B-Instruct-GGUF HF repo, cross-confirmed with HuggingFaceTB official GGUF repo. Ollama shows 1.8GB. Context 8K. Efficient edge/on-device model. |
qwen2.5-0.5b | Qwen2.5 0.5B | Qwen2.5 | 0.494 | false | null | null | null | null | null | null | null | 2024-09-19 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen2.5:0.5b",
"https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Smallest Qwen2.5 dense model. Trained on 18T tokens, supports 29+ languages, 128K context (generates up to 8K). Ideal for on-device and microcontroller-class edge inference. |
qwen2.5-1.5b | Qwen2.5 1.5B | Qwen2.5 | 1.54 | false | null | null | null | null | null | null | null | 2024-09-19 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen2.5:1.5b",
"https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Strong 1.5B edge model from Alibaba. 128K context, 8K generation. Popular for phone and laptop inference with good multilingual and instruction-following capability. |
qwen2.5-3b | Qwen2.5 3B | Qwen2.5 | 3.09 | false | null | null | null | null | null | null | null | 2024-09-19 | null | null | null | Qwen Research License | no | Qwen Research License: non-commercial use only. | null | [
"https://ollama.com/library/qwen2.5:3b",
"https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF",
"https://qwenlm.github.io/blog/qwen2.5/"
] | Mid-range Qwen2.5 dense model. 128K context window, good coding and math performance relative to size. Runs on most 8GB RAM devices at Q4. |
qwen3-0.6b | Qwen3 0.6B | Qwen3 | 0.6 | false | null | null | null | null | null | null | null | 2025-04-29 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3:0.6b",
"https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF",
"https://github.com/QwenLM/Qwen3",
"https://apxml.com/models/qwen3-1-7b",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Smallest Qwen3 dense model. Supports thinking/non-thinking dual mode, 32K native context (extendable to 128K via YaRN). Designed for extreme edge and on-device use. |
qwen3-1.7b | Qwen3 1.7B | Qwen3 | 1.7 | false | null | null | null | null | null | null | null | 2025-04-29 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3:1.7b",
"https://huggingface.co/bartowski/Qwen_Qwen3-1.7B-GGUF",
"https://github.com/QwenLM/Qwen3",
"https://apxml.com/models/qwen3-1-7b",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Dual-mode (thinking/non-thinking) 1.7B dense model. 32K native context (YaRN to 128K). Strong instruction-following and tool-use for its class. |
qwen3-4b | Qwen3 4B | Qwen3 | 4 | false | null | null | null | null | null | null | null | 2025-04-29 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/qwen3:4b",
"https://huggingface.co/Qwen/Qwen3-4B-GGUF",
"https://github.com/QwenLM/Qwen3"
] | Claimed to rival Qwen2.5-72B-Instruct on several benchmarks. 32K native context (YaRN to 128K). Dual thinking/non-thinking mode. Excellent laptop-class edge model. |
gemma-2-2b | Gemma 2 2B | Gemma | 2.61 | false | null | null | null | null | null | null | null | 2024-07-31 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma2:2b",
"https://huggingface.co/bartowski/gemma-2-2b-it-GGUF",
"https://huggingface.co/blog/gemma-july-update",
"https://huggingface.co/blog/gemma2"
] | Google's smallest Gemma 2 model. Uses sliding-window attention (local 4K + global 8K). Strong code and reasoning for its size. Released July 31, 2024 separately from the 9B/27B launch. |
gemma-3-1b | Gemma 3 1B | Gemma | 1 | false | null | null | null | null | null | null | null | 2025-03-12 | null | null | null | Gemma | conditional | Gemma Terms of Use apply to commercial use. | null | [
"https://ollama.com/library/gemma3:1b",
"https://huggingface.co/bartowski/google_gemma-3-1b-it-GGUF",
"https://developers.googleblog.com/en/introducing-gemma3/",
"https://llm-stats.com/models/gemma-3-1b-it",
"https://gorilla.cs.berkeley.edu/leaderboard.html"
] | Google's ultra-small Gemma 3 model. 32K context, multilingual support, strong reasoning for 1B scale. Designed to run on a single GPU or phone-class hardware. |
smollm2-135m | SmolLM2 135M | SmolLM2 | 0.135 | false | null | null | null | null | null | null | null | 2024-11-02 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/smollm2:135m",
"https://huggingface.co/bartowski/SmolLM2-135M-Instruct-GGUF",
"https://venturebeat.com/ai/ai-on-your-smartphone-hugging-faces-smollm2-brings-powerful-models-to-the-palm-of-your-hand"
] | Hugging Face's smallest SLM. 135M params, fits in ~105MB at Q4_K_M. Designed for microcontrollers and on-device inference. Trained on 2T tokens (FineWeb-Edu, DCLM, The Stack). |
smollm2-360m | SmolLM2 360M | SmolLM2 | 0.362 | false | null | null | null | null | null | null | null | 2024-11-02 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/smollm2:360m",
"https://huggingface.co/bartowski/SmolLM2-360M-Instruct-GGUF",
"https://venturebeat.com/ai/ai-on-your-smartphone-hugging-faces-smollm2-brings-powerful-models-to-the-palm-of-your-hand"
] | Mid-tier SmolLM2. 362M params, 271MB at Q4_K_M. Trained on 4T tokens. Suitable for phones and embedded Linux devices. Apache 2.0 license. |
tinyllama-1.1b | TinyLlama 1.1B | TinyLlama | 1.1 | false | null | null | null | null | null | null | null | 2024-01-04 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/tinyllama:1.1b",
"https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF",
"https://github.com/jzhang38/TinyLlama"
] | Community classic 1.1B model pretrained on 3T tokens using Llama-2 architecture. 2048 token context. Very small GGUF footprint (669MB Q4_K_M). Good baseline for edge benchmarking. |
granite-3.1-2b | Granite 3.1 2B | Granite | 2.53 | false | null | null | null | null | null | null | null | 2024-12-18 | null | null | null | Apache-2.0 | yes | null | null | [
"https://ollama.com/library/granite3.1-dense:2b",
"https://huggingface.co/bartowski/granite-3.1-2b-instruct-GGUF",
"https://community.ibm.com/community/user/blogs/nickolus-plowden/2025/01/12/granite-31-delivers-powerful-performance-longer-co",
"https://huggingface.co/ibm-granite/granite-3.1-2b-instruct"
] | IBM's enterprise-grade 2B dense SLM. 128K context (extended from 4K via progressive RoPE training on ~500B tokens). Optimized for tool-use and RAG workloads. Apache 2.0. |
phi-3.5-mini | Phi-3.5-mini 3.8B | Phi-3.5 | 3.82 | false | null | null | null | null | null | null | null | 2024-08-23 | null | null | null | MIT | yes | null | null | [
"https://ollama.com/library/phi3.5:3.8b",
"https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF",
"https://azure.microsoft.com/en-us/blog/new-models-added-to-the-phi-3-family-available-on-microsoft-azure/",
"https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/discover-the-new-multi-lingual-h... | Microsoft's 3.8B long-context SLM. 128K context, multilingual (20+ languages). Trained on 3.4T tokens. Strong performance vs similarly-sized models on reasoning and document QA. |
sarvam-m-24b | Sarvam-M 24B | Sarvam | 24 | false | null | null | null | null | null | null | null | 2025-05 | null | null | text | Apache-2.0 | yes | null | null | [
"https://huggingface.co/sarvamai/sarvam-m",
"https://huggingface.co/lmstudio-community/sarvam-m-GGUF",
"https://huggingface.co/sarvamai/sarvam-m-q8-gguf"
] | Dense 24B fine-tuned from Mistral-Small-3.1-24B-Base. Hybrid thinking mode. Q4_K_M 14.3GB and Q8_0 25.1GB confirmed from two independent GGUF repos (lmstudio-community, Mungert) plus the official sarvamai Q8 repo. Context 32K from config.json. |
sarvam-1-2b | Sarvam-1 2B | Sarvam | 2 | false | null | null | null | null | null | null | null | 2024-10 | null | null | text | Sarvam non-commercial | no | Non-commercial use only per the Sarvam license. Check the current HuggingFace model card for updates. | null | [
"https://huggingface.co/sarvamai/sarvam-1",
"https://huggingface.co/bartowski/sarvam-1-GGUF"
] | 2B dense model trained from scratch, optimized for 10 Indic languages plus English. Released 2024-10-24 under the Sarvam non-commercial license (not Apache). GGUF sizes from bartowski: Q4_K_M 1.55GB, Q8_0 2.69GB. Context 8K from the model card. |
sarvam-30b | Sarvam-30B | Sarvam | 30 | true | null | null | null | null | null | null | null | 2026-03 | null | null | text | Apache-2.0 | yes | null | null | [
"https://huggingface.co/sarvamai/sarvam-30b",
"https://huggingface.co/sarvamai/sarvam-30b-gguf",
"https://www.sarvam.ai/blogs/sarvam-30b-105b"
] | MoE with 128 sparse experts, top-6 routing, 2.4B active params. Released 2026-03 under Apache 2.0. Q4_K_M 19.6GB confirmed by summing the 6 shards in the official GGUF repo. No official Q8_0 GGUF released. |
End of preview.
localmodel.run catalog
The dataset behind localmodel.run: sourced memory requirements for 153 local AI models (125 text LLMs plus image, video and audio models) checked against 40 devices (Apple Silicon Macs, NVIDIA/AMD GPUs, unified-memory APUs, iPhones, Android phones, CPU-only laptops).
Every row carries a sources[] array pointing at its primary source: the
HuggingFace GGUF repo, the Ollama tag, or the vendor spec sheet. GGUF sizes are
measured from the actual files, never estimated from parameter counts.
Files
| File | Contents |
|---|---|
models.json |
125 text LLMs: params, measured GGUF sizes (Q4_K_M, Q8_0), context length, MoE active params, Ollama tag, benchmark scores (Aider polyglot, BFCL, LMArena Elo) where sourced |
devices.json |
40 devices: usable memory, memory bandwidth, MSRP and TDP, sourced to Apple, NVIDIA and AMD spec pages |
image-models.json, video-models.json, audio-models.json |
28 diffusion and audio models with sourced peak-VRAM or peak-memory anchors |
tools.json |
per-platform runtime recommendations |
The memory math
The site computes, per model and device pair:
total = weights + KV cache + overhead
Weights are the measured GGUF size at the chosen quant. The KV cache grows with context length. Usable memory is device-specific: the same 16 GB sticker gives a discrete GPU about 15 GB, an iPad about 12 GB and a Mac about 10.5 GB. The full formula, constants and thresholds are public: localmodel.run/methodology. The figures are estimates computed from these sourced inputs.
Attribution (CC BY 4.0)
Reuse the data in your app, benchmark, paper or README with a link:
Data: [localmodel.run](https://localmodel.run) (CC BY 4.0)
Related
- Website: localmodel.run, one page per model and device pair, about 6,400 pages
- Live demo: Can I Run It Locally? (HF Space)
- Source and contributions: github.com/ansumanshah/localmodel.run, with issue forms for adding a model or reporting a wrong number
- Freshness: a weekly cron re-pulls GGUF sizes from the Ollama registry and the HuggingFace Hub and flags drift over 15%
- Downloads last month
- 32