Instructions to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX"
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX" } ] } } }Run Pi
# Start Pi in your project directory: pi
- MLX LM
How to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX", "messages": [ {"role": "user", "content": "Hello"} ] }' - Hermes Agent
How to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
Qwen3.8-27B-UD3-Q3_K_XL-MLX
Native MLX affine re-encoding of the official Unsloth Dynamic v3.0 Q3_K_XL per-tensor allocation for Qwen3.8-27B.
This is a 13.63 GB, text-only MLX model with 3.991 effective bits per quantized weight. It retains the model's inline MTP head and is intended for mlx-serve on Apple Silicon.
Before downloading: use this model with
mlx-serve. Stockmlx-lmis not a supported runtime for these published Qwen3.8 MTP layouts.
What this model is
The upstream GGUF determines the bit width of each tensor. maccelerate reads that allocation and quantizes the corresponding clean bf16 Qwen weights onto MLX's affine format at the same width.
This model has the same per-tensor bit-width allocation as the source GGUF, re-encoded for MLX. It is not a lossless GGUF conversion and does not claim numerical parity with llama.cpp's K-quant or IQ codebooks.
| Claim | Status |
|---|---|
| Same upstream per-tensor allocation | Yes |
| Quantized from clean bf16 weights | Yes |
| Exact numerical parity with the GGUF | No |
| MTP head | Retained inline; one MTP layer |
| Vision tower | Omitted; this is a text-only release |
Allocation
| MLX width | Tensors |
|---|---|
| 8-bit | 98 |
| 6-bit | 7 |
| 5-bit | 26 |
| 4-bit | 194 |
| 3-bit | 157 |
| 2-bit | 24 |
| Total | 506 |
Widths are not a house preset: each tensor retains the width selected by the upstream Dynamic allocator. The allocation and structural checks for this exact build are recorded in the included manifest.json.
Serving
Apple Silicon requirements
- Apple Silicon Mac running macOS.
- Allow at least 14 GB of disk for the model files, plus working space for the download.
- The model files occupy 13.63 GB. Actual unified-memory needs also depend on the KV cache, context length, and concurrent requests; no minimum-memory configuration is claimed by this card.
Run it with mlx-serve
Install and run mlx-serve according to its README, download this repository locally, then point it at that directory:
mlx-serve --model /path/to/Qwen3.8-27B-UD3-Q3_K_XL-MLX --kv-quant 8
--kv-quant 8 is a reference configuration, not a memory guarantee for every context length or workload. The pack is mixed-width; MTP is present in the checkpoint, so use a runtime configuration that enables speculative decoding if desired.
Runtime compatibility
Use mlx-serve for this release. Stock mlx-lm is not supported for the published Qwen3.8 MTP layout; loading can succeed without providing a supported result.
Reproducibility and generation code
This checkpoint was generated with maccelerate, the open-source converter that reads a Dynamic GGUF's per-tensor allocation and re-encodes the matching clean bf16 model for MLX.
Provenance
| Item | Value |
|---|---|
| Source GGUF | unsloth/Qwen3.8-27B-GGUF, Qwen3.8-27B-UD-Q3_K_XL.gguf |
| Source GGUF revision | Not recorded in the published manifest |
| Source GGUF SHA-256 | 8c2a45ff85e7674ca185ec8eb6cdeab0e617ed9d8018caed0b64380eb2a67a5e |
| bf16 source | Qwen/Qwen3.8-27B |
| bf16 source revision | Not recorded in the published manifest |
| imatrix | imatrix_unsloth.gguf |
| Converter | maccelerate (revision not recorded in the published manifest) |
| Output | 7 SafeTensors shards; 13.63 GB |
The included manifest.json records the source file and hash, allocation, structural validation, and shard hashes for this exact build.
License
The model weights, source GGUF, and conversion code are licensed under Apache-2.0. See the included LICENSE and the upstream model cards for attribution and terms.
- Downloads last month
- 191
4-bit
Model tree for maccelerate/Qwen3.8-27B-UD3-Q3_K_XL-MLX
Base model
Qwen/Qwen3.8-27B