rigbench-lab/rigbench-local-llm-speeds

RigBench local LLM speeds

How fast open language models run on a given GPU, Mac or APU, in tokens per second. One row per published benchmark run: model, quantization, hardware, backend, generation speed and, where reported, prompt speed. Every row links to the post or page the numbers came from.

This is the data behind rigbench.pages.dev. Snapshot: 2026-10-22 (UTC). 660 rows, 41 models, 210 hardware entries.

Results from LocalScore, which the site also shows, are left out for now, pending permission to republish them.

Where the numbers come from

RigBench does not run these benchmarks itself. It collects runs that people have already published, copies the numbers as written, and records the source URL, the llama.cpp commit and the contributor where given. No number is edited or estimated. The site's fit and speed estimates are not in this dataset.

Source	source_id	Rows
llama.cpp Vulkan scoreboard (#10879)	llamacpp-10879	237
llama.cpp CUDA scoreboard (#15013)	llamacpp-15013	112
llama.cpp Apple Silicon thread (#4167)	llamacpp-4167	96
Hand-checked individual posts, READMEs and model cards (26 pages)	curated	82
llama.cpp ROCm scoreboard (#15021)	llamacpp-15021	46
llama.cpp discussion #15396	llamacpp-15396	34
PMZFX intel-arc-pro-b70-benchmarks (MIT)	pmzfx-b70	16
llama.cpp benches/ (maintainer runs)	llamacpp-benches	14
ml-explore/mlx-lm BENCHMARKS.md (MIT)	mlx-lm-benchmarks	11
john-rocky apple-silicon-llm-bench (MIT)	apple-silicon-llm-bench	9
llama.cpp discussion #16578	llamacpp-16578	3

• llama.cpp discussions are the community scoreboards for Apple Silicon (#4167), CUDA (#15013), Vulkan (#10879) and ROCm (#15021), plus the gpt-oss guide (#15396) and the DGX Spark thread (#16578). These are llama-bench runs.
• pmzfx-b70 rows are llama-bench SYCL runs on one Intel Arc Pro B70, from the author's published JSON files (default config only; multi-GPU and long-context variants left out).
• mlx-lm-benchmarks and apple-silicon-llm-bench rows ran Apple's MLX framework (not llama.cpp) on an M4 Max with 40 GPU cores. Their quant labels start with MLX-, gguf_file_bytes is empty, and apple-silicon-llm-bench reports a warm-median short-chat decode speed, not tg128.
• curated rows were copied by hand from individual GitHub comments, READMEs and model cards for newer models. The evidence column holds the exact source lines for each one.

Checks applied

Generating a token means reading every active weight from memory once, so a chip cannot generate faster than its memory bandwidth divided by the bytes read per token. Rows more than 10% above that ceiling are dropped (for MoE models, only the active experts count). Rejected rows are listed on the site at https://rigbench.pages.dev/data/rejected.json and are not in this dataset. Source hardware names are mapped to one RigBench hardware entry, and that entry's spec values (VRAM, bandwidth) come from a cited spec sheet; see https://rigbench.pages.dev/specs/.

Columns

Column	Type	Description
row_id	string	Unique row key in this dataset. Equals rigbench_id, plus a -2, -3 ... suffix when one source post holds several runs that share an id.
rigbench_id	string	Row id used on rigbench.pages.dev (hash of source and settings; not unique when a post lists several runs).
model	string	RigBench model slug.
model_name	string	Model display name.
model_params_b	float64	Total parameters, billions (from the model's Hugging Face GGUF repo).
model_is_moe	bool	Mixture-of-experts model.
quant	string	GGUF quantization as stated by the source (e.g. Q4_0, Q4_K_M, MXFP4, UD-Q4_K_XL).
gguf_file_bytes	int64	Size of that model+quant GGUF file on Hugging Face, where RigBench has it. Reference value, not measured. Empty for custom quants.
hardware	string	RigBench hardware slug (one GPU, Mac or APU).
hardware_name	string	Hardware display name.
hardware_vendor	string	NVIDIA, AMD, Apple, Intel or Other.
hardware_kind	string	gpu, laptop, workstation, datacenter, mac or apu.
hardware_memory_gb	float64	VRAM in GB from the spec sheet. Empty for Macs (unified memory varies by config).
hardware_bandwidth_gbs	float64	Peak memory bandwidth in GB/s from the spec sheet (source on rigbench.pages.dev/specs/).
hardware_raw	string	Hardware name exactly as the source wrote it.
backend	string	Runtime backend: CUDA, Vulkan, Metal, ROCm, etc. Empty when the source did not say.
flash_attn	bool	Flash attention on/off. Empty when not reported.
method	string	How the run was timed: llama-bench, llama-server, mlx_lm.benchmark or apple-silicon-llm-bench.
tg_tps	float64	Generation speed, tokens per second. Always present.
tg_test	string	Which generation test: tg128 = 128 generated tokens (llama-bench); tg32 etc. = other lengths; free text for llama-server runs.
pp_tps	float64	Prompt processing speed, tokens per second. Empty when not reported.
pp_test	string	Which prompt test: pp512 = 512-token prompt (llama-bench), pp2048, etc.
partial_offload	bool	True when the model file is larger than the card's VRAM, so part of it ran from system RAM.
llamacpp_commit	string	llama.cpp build/commit when the source gave one (for MLX rows: the mlx-lm version).
contributor	string	Who ran the benchmark, when the source names them (GitHub handle).
source_id	string	Short source key (llamacpp-15013, curated, ...).
source_name	string	Human-readable source name.
source_url	string	Link to the post or page the numbers were copied from.
notes	string	RigBench's notes on settings or caveats for this row.
evidence	string	For hand-checked rows: the exact source lines the numbers were copied from, joined with '
first_seen	string	Date (YYYY-MM-DD) the row first appeared in RigBench.
retrieved_at	string	UTC time the source was last fetched.

Contents

• 660 rows. Methods: llama-bench 585, llama-server 55, mlx_lm.benchmark 11, apple-silicon-llm-bench 9.
• Hardware vendors: NVIDIA 294, AMD 173, Apple 146, Intel 45, Other 2.
• 568 rows use the standard llama-bench tg128 test; the rest are other llama-bench lengths, llama-server or MLX timings described in tg_test and notes. The tg128 count includes mlx_lm.benchmark rows, which are MLX runs.

Model	Rows
Llama 2 7B	491
Qwen3.8 27B	53
gpt-oss 20B	30
gpt-oss 120B	11
Qwen3 4B Instruct 2507	7
Gemma 4 E4B	6
Qwen3.6 35B-A3B	6
Gemma 4 26B-A4B	4
Qwen3 30B-A3B Instruct 2507	4
Qwen3.5 27B	4
Qwen3.8 Flash Next	4
Gemma 4 31B	3
GLM-5.3-Flash	3
Gemma 3 4B IT (QAT)	2
GLM-4.7-Flash	2
Hy3	2
Qwen2.5 Coder 7B	2
Qwen3.5 9B	2
Qwen3 Coder 30B-A3B Instruct	2
DeepSeek R1 Distill Qwen 1.5B	1
Devstral Small 2 24B	1
Gemma 3 1B IT	1
Inkling Small	1
Kimi K2.7 Code	1
Llama 3.1 8B Instruct	1
Llama 3.2 3B Instruct	1
MiMo-V2.6-Flash	1
Mistral Small 3.2 24B Instruct 2506	1
Muse Spark Open	1
Nemotron 3 Nano 4B	1
Nemotron 3 Super 120B-A12B	1
Nex-N2-Pro	1
Phi-4 14B	1
Phi-4 mini instruct	1
Qwen2.5 1.5B	1
Qwen3 0.6B	1
Qwen3 1.7B	1
Qwen3 4B	1
Qwen3.5 35B-A3B	1
SmolLM3 3B	1
TinySwallow 1.5B Instruct	1

Using it

from datasets import load_dataset
ds = load_dataset("rigbench-lab/rigbench-local-llm-speeds", split="train")
df = ds.to_pandas()
# compare like with like: same model, quant and test
q = df[(df.model == "llama-2-7b") & (df.quant == "Q4_0") & (df.tg_test == "tg128")]
print(q.sort_values("tg_tps", ascending=False)[["hardware_name", "backend", "tg_tps", "pp_tps"]].head(20))

Only compare rows with the same model, quant, tg_test and ideally backend. Different tests measure different things.

Limits

• Not a controlled benchmark. Runs come from many people, machines, drivers, llama.cpp builds and settings. Two runs on the same card can differ noticeably.
• Coverage is uneven. 491 of 660 rows are Llama 2 7B (the llama.cpp scoreboard model). Most newer models have only one to a few rows.
• Mixed test types. tg128, other llama-bench lengths and llama-server decode timings are not directly comparable.
• Missing fields. Some sources do not report prompt speed, flash attention, backend, commit or contributor; those cells are empty.
• Repeated ids. 18 rows share a rigbench_id with another row because one post listed several runs whose differing settings were not captured. Use row_id as the key.
• Partial offload. 7 rows ran with part of the model in system RAM (partial_offload = true). They are much slower than a fully loaded card.
• Hardware mapping. Source names are matched to one RigBench entry; laptop GPUs and some Mac configurations vary more than the single spec value suggests.

Updates

The site rebuilds its data daily from the sources above. This dataset is a snapshot of that data and is re-exported with export.py; the snapshot date is at the top of this card. The live CSV is always at https://rigbench.pages.dev/data/benchmarks.csv.

Related data not included

The open-weight model rankings from Artificial Analysis are shown on https://rigbench.pages.dev/models/ but are their data, not ours, and are not re-hosted here. Use Artificial Analysis directly for those.

License and credit

The compilation (normalization, hardware and model mapping, checks, notes) is released under CC BY 4.0. The measurements themselves were published by their original authors: keep the source_url when you reuse a row, and credit the original source where you can. Posts in the llama.cpp GitHub discussions belong to their authors (llama.cpp's code is MIT, which does not cover them). If you need clear terms for a commercial use, check with the original source.

Rows from these sources are redistributed under the MIT License; the copyright notices apply to those rows:

• pmzfx-b70: PMZFX/intel-arc-pro-b70-benchmarks, Copyright (c) 2026 Kato Papairo (MIT License)
• mlx-lm-benchmarks: ml-explore/mlx-lm BENCHMARKS.md, Copyright (c) 2023 Apple Inc. (MIT License)
• apple-silicon-llm-bench: john-rocky/apple-silicon-llm-bench, Copyright (c) 2026 john-rocky and ios-llm-benchmark contributors (MIT License)

To credit this dataset:

RigBench local LLM speeds, snapshot 2026-10-22. https://huggingface.co/datasets/rigbench-lab/rigbench-local-llm-speeds, compiled from published benchmark runs; see https://rigbench.pages.dev

Corrections: every row links to its source. If a number does not match the source, use the correction form at https://rigbench.pages.dev/submit/#correction.