whatmodel.sh / Qwen3 8B

Qwen3 8B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Qwen3 8B is an open-weights language model from Alibaba, released April 2025. Its published size is 8.19B parameters and a 33k-token context window. Licence: apache-2.0.

55.7±0.6#251 of 719 LLMsfrom 46 benchmark results across 20 sources

$ how it compares

The LLMs ranked either side of Qwen3 8B, out of 828 with a score, then more from Alibaba.

#modelscorewhereRAMreleased
300Mistral Small 3.2 24B Mistral56±1api $0.11/M +local~16GB2025-06
301DeepSeek-V2.5 (1210) DeepSeek56±1local~82GB2024-12
302Qwen3 8B Alibaba56±1api $0.31/M +local~6.1GB2025-04
303DeepSeek R1 Distill Qwen 14B DeepSeek56±1local~10GB2025-01
304Ministral 3 14B Mistral56±1api $0.20/M +local~9.1GB2025-12
more from Alibaba
18Qwen3.8 Max (2.4T-A95B) Alibaba78±1api $3.00/M +local~380GB2026-08
34Qwen3.8 Max (0902) Alibaba76±1api $3.00/M2026-09
37Qwen3.8-Flash-Next Alibaba75±1local~59GB2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #188 of 300 · 19% of its score · 6 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Mellum2 12B-A2.5B Instruct here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
τ²-bench Telecom +219.1% ✎█··██··██··█ beats 0/5very hard · avg 10%▇▇▇▇▇ 12%sharp · several runs
Terminal-Bench Hard2.1% ✎▂▅▆▂▂▆▂▂█▅▅▂ beats 1/27very hard · avg 1.4%▇ 2.2%sharp · too hard for it
BFCL v3 +166.3% ✎▁▁▁▁▁▃▅█▅▄▁▁ beats 100/114medium · avg 59%▇ 2.0%blunt · at its level · several runs
BFCL V4 (December 2025 archive)42.6% ✎▁▁▁▂██▅▅▆▂▂▂ beats 49/77hard · avg 32%▇ 1.5%at its level
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)33.2% ✎▂·▁▃··▇▇█▂·▁ beats 3/25hard · avg 29%▇ 0.8%blunt · at its level
BFCL (version not stated)60.2% ✎▄▆▄▂·█▆·▄··▂ beats 4/17easy · avg 66%▇ 0.7%blunt · at its level
coding▁▁▁▃▆██▇▆▄▃▁ #245 of 458 · 5.7% of its score · 2 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails QwQ-32B-Preview, InternLM2.5 20B Chat, Mistral Small Instruct 22B (2409) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench v6 (window unspecified) lab +151.0% ✎▃▅█▅▅▃▅···▃█ beats 8/16medium · avg 35%▇▇ 5.2%at its level · several runs
SciCode main problems6.7% ✎▄··▄····█▄·▄ beats 3/5very hard · avg 3.8%▇ 0.5%too hard for it
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #177 of 393 * · 1.8% of its score · 1 test

Beats Seed-OSS 36B Instruct, Nemotron-Cascade-14B-Thinking, Qwen3-VL 32B Instruct here, though they rank higher overall. Trails LFM2.5 1.2B Instruct, LFM2.5 1.2B Thinking, Gemma 3 12B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
IFEval (aggregation unspecified) lab85.0% ✎▃··▃···██▅▃▃ beats 9/11easy · avg 77%▇ 1.8%typical
math▁▂▃▅▆▅▄█▄▁▁▁ #170 of 398 · 27% of its score · 7 tests

Beats MiniMax-M2.1, Mistral Medium 3.5 (128B), GPT-4.1 here, though they rank higher overall. Trails OpenThinker3 7B, Olmo 3 7B Think, OpenThinker2 32B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
AIME 2025 lab +167.3% ✎▃·▃▃█··▃▅▃·█ beats 7/12medium · avg 53%▇▇▇ 7.1%sharp · at its level · several runs
AIME 202477.9% ✎█▄···█·····▄ beats 1/5medium · avg 62%▇▇ 4.7%sharp
HMMT 2025 (edition unspecified)41.0% ✎▄··▄▄·▄····█ beats 0/5hard · avg 29%▇▇ 4.1%sharp · at its level
OMEGA (math generalisation) +143.4% ✎▄······█▄▄·█ beats 6/6hard · avg 30%▇▇ 3.9%at its level · several runs
MATH (subset unspecified) +195.1% ✎▃·····▃···▅█ beats 4/6very easy · avg 90%▇ 2.8%maxed out · several runs
OTIS mock AIME 2024–202556.1% ✎▃▃▄▃▃▃▇▇█▃▂▅ beats 85/193very hard · avg 8.5%▇ 2.4%sharp · noisy
MATH-50097.4% ✎▄·▄·····█··█ beats 4/5very easy · avg 92%▇ 2.1%maxed out
other tests21 not in a task · 46% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond +156.8% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 84/208hard · avg 40%▇▇▇▇ 10%sharp · at its level · several runs
MMLU-Pro lab +174.6% ✎▄·▄·▄▄▄▄█▄██ beats 7/12medium · avg 63%▇▇▇ 6.8%at its level · several runs
GPQA (subset unspecified) lab62.0% ✎▂····▂·▂▂█▆█ beats 5/14hard · avg 49%▇▇ 5.6%sharp · at its level
MMLU-Redux lab87.5% ✎▃··▃·▃·█▃▅▃▅ beats 6/11easy · avg 79%▇ 3.6%typical
SuperGPQA lab47.6% ✎▄··▄·▄▄▄█▄██ beats 6/11medium · avg 37%▇ 2.7%at its level
Medmarks-V weighted mean win rate (61 models, 71 configurations)50.2% ✎▂▃▅▂▆▃█▅█▅▃▆ beats 17/34medium · avg 42%▇ 2.4%blunt · at its level
EQ-Bench 3 (emotional intelligence, judge Elo)1083.3 ✎▁·▂▁▄▅▅▃█▄▁▁ beats 11/62avg 1044▇ 2.0%typical
CPTUB average (SpeakLeash, 0-5 judge scale)61.2% ✎▁▂▂▂▃█▃▆▅▄▂▃ beats 35/95easy · avg 66%▇ 1.9%at its level
13 more tests
testresultvs the others in its rundifficultycountswhy
Humanity's Last Exam4.2% ✎█··██···█·██ beats 0/5very hard · avg 3.8%▇ 1.8%too hard for it
AraGen-v3 3C3H score (Inception, Claude Sonnet 4 judge)26.6% ✎▁▁·▁▂▂▄█▄▅▄▄ beats 23/66hard · avg 31%▇ 1.7%at its level
BRIDGE clinical text, zero-shot average (87 tasks)40.0% ✎▁▁▁▁▁▂▂▃▄▆█▃ beats 75/97hard · avg 32%▇ 1.7%blunt · at its level
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items22.3% ✎▁▁▂▁▃▄█▆▄▅▂▃ beats 6/39very hard · avg 22%▇ 1.5%blunt
Open Persian LLM Leaderboard v2 average (21 tasks)55.0% ✎▂▄▂▃█▄▃▆▃▅▂▂ beats 41/81medium · avg 58%▇ 1.3%blunt · at its level
IFBench (aggregation unspecified)35.3% ✎███···██···█ beats 1/5hard · avg 35%▇ 0.6%at its level
DiagnosticIQ industrial maintenance MCQ (macro accuracy, 4 options)46.2% ✎▁··▃▁▁▁▅█▅▁▅ beats 1/24medium · avg 52%▇ 0.6%blunt · at its level
Open PT-LLM Leaderboard average (9 Portuguese tasks)78.3% ✎▃▁▁▁▁▁▁▁▂█▅▁ beats 243/276medium · avg 73%▇ 0.6%at its level · noisy
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy75.5% ✎▃·▁▁▂▁▂▅█▇▄▁ beats 60/100medium · avg 68%▇ 0.5%at its level · noisy
Chess Puzzles (Epoch)5.0% ✎█▁▁▁▁▂▁▄▂▂▁▁ beats 21/135very hard · avg 5.9%▇ 0.3%sharp · too hard for it · noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy89.8% ✎▁▁▁▁▁▁▁▂▄█▄▃ beats 115/146easy · avg 83%▇ 0.3%noisy
Vectara hallucination leaderboard, 2025 dataset (HHEM-2.3, factual consistency rate)97.0% ✎▁▁▂▃▄▄█▆▃▂▂▁ beats 102/140very easy · avg 94%▇ 0.2%blunt · maxed out · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)95.2% ✎▁▂▁▂▅█▂▃▄▃▁▁ beats 92/98very easy · avg 90%▇ 0.0%blunt

$ where to run it

Through an API

api · $0.31/M · $0.18 in / $0.7 out per 1M tokens · 33k context · www.alibabacloud.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

version

  1. Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
    macOS / Linux
    curl -LsSf https://llama.app/install.sh | sh
    Windows
    irm https://llama.app/install.ps1 | iex

    other ways to install ↗

  2. Run Qwen3 8B · needs ~5 GB · from huggingface.co/Qwen (official)
    chat in the terminal
    llama cli -hf Qwen/Qwen3-8B-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf Qwen/Qwen3-8B-GGUF:Q4_K_M

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run Qwen3 8B · needs ~5 GB · from huggingface.co/Qwen (official)

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/Qwen/Qwen3-8B-GGUF@Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

  1. Install Ollama

    download it (macOS / Windows) ↗

    Linux
    curl -fsSL https://ollama.com/install.sh | sh
  2. Run Qwen3 8B · needs ~5.41 GB · from ollama.com
    chat in the terminal
    ollama run qwen3:8b

MLX (Mac)

version

  1. Install MLX (Mac) · Apple silicon Macs only, text only.
    with uv
    uv tool install mlx-lm
    or
    pip install mlx-lm
  2. Run Qwen3 8B · needs ~4.4 GB, Mac only · from huggingface.co/Qwen (official)
    chat in the terminal
    mlx_lm.chat --model Qwen/Qwen3-8B-MLX-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model Qwen/Qwen3-8B-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Falcon-H1R: Pushing the Reaso… 33%, Qwen (Alibaba) vendor-reporte… 25%, Epoch AI — internal evaluatio… 8% and 17 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling12▇▇▇▇▇▇▇▇▇▇ 33%CC BY 4.0
Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card) · lab-reported8▇▇▇▇▇▇▇▇ 25%Factual vendor-reported score (not licensed content)
Epoch AI — internal evaluations3▇▇ 8%CC BY 4.0
MUA-RL: Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use (arXiv:2508.18669v1)1▇▇ 6%CC-BY-4.0
allenai/Olmo-3-7B-Think model card2▇ 5%Apache-2.0
Synthesize and Reward: Reinforcement Learning for Multi-Step Tool Use in Live Environments (arXiv:2606.03892v2)1▇ 3%CC-BY-4.0
allenai/Olmo-3-7B-Instruct model card3▇ 2%Apache-2.0
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models1▇ 2%CC BY 4.0
Berkeley Function Calling Leaderboard result archive2▇ 2%Apache-2.0
EQ-Bench 3 canonical leaderboard results1▇ 2%MIT
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)1▇ 2%MIT
Inception Arabic/Hindi leaderboards (AraGen-v3, HindiGen-v1, Arabic IFEval)1▇ 2%Apache-2.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)1▇ 2%MIT
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)1▇ 2%CC BY 4.0
Open Persian LLM Leaderboard v2 (Part DP AI and AUT NLP Lab)1▇ 1%Apache-2.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)1▇ 1%CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard2▇ 1%MIT
DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules (arXiv:2605.08614v1)1▇ 1%CC BY 4.0
Open Portuguese LLM Leaderboard (archived)1▇ 1%Apache-2.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)2▇ 0%Apache-2.0
prior · pull toward the middle▇ 0%
weighted average= 55.7
$ published speed measurements · 4

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

NVIDIA GeForce RTX 5090 (32GB; host CPU/RAM not stated)NVFP4 weights, FP16 activations151.4 tok/sTensorRT-LLM v1.1.0, PyTorch backend (default) · Longest prompt variant (~496 input tokens); fixed 256 output tokens, greedy decoding, fixed seed; two untimed warm-up generations; mean of three timed runs; generation throughput measured strictly during the decode phase; Ubuntu 24.04

Table 4 'backend dichotomy' row.

source ↗
NVIDIA GeForce RTX 5090 (32GB; host CPU/RAM not stated)NVFP4 weights, FP16 activations93.1 tok/sTensorRT-LLM v1.1.0, C++ engine (legacy) · Longest prompt variant (~496 input tokens); fixed 256 output tokens, greedy decoding, fixed seed; two untimed warm-up generations; mean of three timed runs; generation throughput measured strictly during the decode phase; Ubuntu 24.04

Table 4 'backend dichotomy' row.

source ↗
NVIDIA GeForce RTX 5090 (32GB; host CPU/RAM not stated)GGUF BF1692.2 tok/sllama.cpp (version not stated) · Longest prompt variant (~496 input tokens); fixed 256 output tokens, greedy decoding, fixed seed; two untimed warm-up generations; mean of three timed runs; generation throughput measured strictly during the decode phase; Ubuntu 24.04

Table 4 'backend dichotomy' row.

source ↗
Apple M2 Pro (16-core GPU), 16 GBQ4_K_M21.05 tok/sollama 0.30.8 · whatmodel bench.sh v1: fixed prompt (~168 tokens), 256 generated tokens, temperature 0, median of 3 runs

Community submission, reviewed 2026-09-25. 100% of the model in GPU memory.

source ↗
$ quantization evidence · 129 measurements

How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent.

precisiontestfull → quantizedkept
AutoRound 2-bit (group size 32; 512 calibration samples, seqlen 4096, 512 iterations)WMT24++ X->English COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.782 → 0.62880.4%source ↗
AutoRound 2-bit (group size 32; 512 calibration samples, seqlen 4096, 512 iterations)WMT24++ English->X COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.702 → 0.43261.6%source ↗
AutoRound 4-bit (group size 128; 512 calibration samples, seqlen 4096, 512 iterations)WMT24++ X->English COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.782 → 0.77399.0%source ↗
AutoRound 4-bit (group size 128; 512 calibration samples, seqlen 4096, 512 iterations)WMT24++ English->X COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.702 → 0.68998.1%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6] iters 200' (actual 3.25 bits/weight)Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.61396.9%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6] iters 200' (actual 3.25 bits/weight)MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.69294.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6] iters 200' (actual 3.25 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.787.7%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6] iters 200' (actual 3.25 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.39986.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6]' (actual 3.24 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.69887.5%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6]' (actual 3.24 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.36980.2%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6]' (actual 3.24 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.60796.0%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4,q6]' (actual 3.24 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.6994.5%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4] lm_head Q6K' (actual 3.25 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.71489.5%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4] lm_head Q6K' (actual 3.25 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.34875.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4] lm_head Q6K' (actual 3.25 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.60996.3%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4] lm_head Q6K' (actual 3.25 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.69394.9%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4]' (actual 3.23 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.70688.5%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4]' (actual 3.23 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.36479.1%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4]' (actual 3.23 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.61597.3%source ↗
AutoRound AutoScheme mixed-precision GGUF '2.5b mixed[q2,q3,q4]' (actual 3.23 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.69495.1%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6] iters 200' (actual 4.19 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.505109.9%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6] iters 200' (actual 4.19 bits/weight)Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.63299.9%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6] iters 200' (actual 4.19 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.78898.7%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6] iters 200' (actual 4.19 bits/weight)MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.71898.4%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6]' (actual 4.19 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.814102.0%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6]' (actual 4.19 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.44496.7%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6]' (actual 4.19 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.62598.7%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4,q6]' (actual 4.19 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.71998.5%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4] lm_head Q6K' (actual 4.19 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.78898.7%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4] lm_head Q6K' (actual 4.19 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.4597.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4] lm_head Q6K' (actual 4.19 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.62598.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4] lm_head Q6K' (actual 4.19 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.71898.3%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4]' (actual 4.17 bits/weight)MATH-500 (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.798 → 0.816102.3%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4]' (actual 4.17 bits/weight)GPQA Diamond (EvalScope via llama-server, --reasoning off, --predict 2048, ctx 65536, --parallel 16)0.46 → 0.4597.8%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4]' (actual 4.17 bits/weight); no AutoRound tuning per line 35Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.62899.3%source ↗
AutoRound AutoScheme mixed-precision GGUF '3.5b mixed[q2,q3,q4]' (actual 4.17 bits/weight); no AutoRound tuning per line 35MMLU (lm-eval default mmlu task; lm_eval --model hf with gguf_file)0.73 → 0.72399.0%source ↗
AWQ 4-bit (AutoAWQ, zero point, group size 128)WMT24++ X->English COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.782 → 0.77499.0%source ↗
AWQ 4-bit (AutoAWQ, zero point, group size 128)WMT24++ English->X COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.702 → 0.6998.3%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)MMLU-Redux, non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.795 → 0.79199.5%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)MMLU-Redux, thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.875 → 0.86498.7%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)LiveBench 2024-11-25, thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.671 → 0.65597.6%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)GPQA (subset not stated in card), thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.62 → 0.5995.2%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)AIME 2024, thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.76 → 0.71393.8%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)LiveBench 2024-11-25, non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.535 → 0.48991.4%source ↗
AWQ INT4 (w4, group_size 128, zero_point, GEMM; AutoAWQ; checkpoint Qwen/Qwen3-8B-AWQ)GPQA (subset not stated in card), non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)0.393 → 0.35991.3%source ↗
bitsandbytes NF4 (Transformers integration, nested/double quantization)WMT24++ X->English COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.782 → 0.77699.3%source ↗
bitsandbytes NF4 (Transformers integration, nested/double quantization)WMT24++ English->X COMET (wmt22-comet-da), unweighted mean over 6 language pairs (ja_JP, fr_FR, pl_PL, bn_IN, ml_IN, zu_ZA); greedy decoding0.702 → 0.69298.6%source ↗
FP8 block (weights FP8 128x128 blocks, activations FP8 dynamic per-group 128; llm-compressor QuantizationModifier)OpenLLM v2 average of 6 tasks (IFEval (Inst Level Strict Acc, 0-shot); BBH (Acc-Norm, 3-shot); Math-Hard (Exact-Match, 4-shot); GPQA (Acc-Norm, 0-shot); MUSR (Acc-Norm, 0-shot); MMLU-Pro (Acc, 5-shot)) as published; lm-evaluation-harness + vLLM (backend vllm, max_model_len 16384, chat template, few-shot as multiturn); per-task scale (raw vs Open LLM Leaderboard-normalized) not stated in card0.289 → 0.291100.6%source ↗
FP8 block (weights FP8 128x128 blocks, activations FP8 dynamic per-group 128; llm-compressor QuantizationModifier)OpenLLM v1 average of 6 tasks (ARC-Challenge (Acc-Norm, 25-shot); GSM8K (Strict-Match, 5-shot); HellaSwag (Acc-Norm, 10-shot); MMLU (Acc, 5-shot); TruthfulQA (MC2, 0-shot); Winogrande (Acc, 5-shot)) as published; lm-evaluation-harness + vLLM (backend vllm, max_model_len 16384, add_bos_token)0.721 → 0.721100.0%source ↗
FP8 block (weights FP8 128x128 blocks, activations FP8 dynamic per-group 128; llm-compressor QuantizationModifier)GSM8K — published row 'GSM8K (Strict-Match, 5-shot)' (OpenLLM v1 section); lm-evaluation-harness + vLLM (backend vllm, max_model_len 16384, add_bos_token)0.879 → 0.87899.8%source ↗
FP8 block (weights FP8 128x128 blocks, activations FP8 dynamic per-group 128; llm-compressor QuantizationModifier)MMLU — published row 'MMLU (Acc, 5-shot)' (OpenLLM v1 section); lm-evaluation-harness + vLLM (backend vllm, max_model_len 16384, add_bos_token)0.749 → 0.74799.8%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)GSM8K — published row 'GSM-8K (5-shot, strict-match)' (OpenLLM v1 section); lm-evaluation-harness + vLLM (backend vllm, max_model_len 8192, chat template, few-shot as multiturn)0.76 → 0.805106.0%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)GPQA Diamond — published row 'GPQA diamond' (Reasoning section); lighteval + vLLM, temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 327680.586 → 0.611104.3%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)OpenLLM v2 average of 6 tasks (MMLU-Pro (5-shot); IFEval (0-shot); BBH (3-shot); Math-lvl-5 (4-shot); GPQA (0-shot); MuSR (0-shot)) as published; lm-evaluation-harness + vLLM (backend vllm, max_model_len 16384, chat template, few-shot as multiturn); per-task scale (raw vs Open LLM Leaderboard-normalized) not stated in card0.343 → 0.353103.0%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)AIME 2024 — published row 'AIME 2024' (Reasoning section); lighteval + vLLM, temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 327680.746 → 0.763102.4%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)OpenLLM v1 average of 6 tasks (MMLU (5-shot); ARC Challenge (25-shot); GSM-8K (5-shot, strict-match); Hellaswag (10-shot); Winogrande (5-shot); TruthfulQA (0-shot, mc2)) as published; lm-evaluation-harness + vLLM (backend vllm, max_model_len 8192, chat template, few-shot as multiturn)0.642 → 0.648101.0%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)LiveCodeBench — published row 'LiveCodeBench' (Reasoning section); lighteval + vLLM, temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 327680.563 → 0.566100.6%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)MMLU — published row 'MMLU (5-shot)' (OpenLLM v1 section); lm-evaluation-harness + vLLM (backend vllm, max_model_len 8192, chat template, few-shot as multiturn)0.72 → 0.723100.5%source ↗
FP8 dynamic (weights FP8 per-channel static, activations FP8 per-token dynamic; llm-compressor QuantizationModifier)AIME 2025 — published row 'AIME 2025' (Reasoning section); lighteval + vLLM, temperature 0.6, top_p 0.95, top_k 20, max_new_tokens 327680.652 → 0.63797.8%source ↗
GGUF Q2_K (bartowski/Qwen3-8B-GGUF file as labelled; actual 3.20 bits/weight)Average accuracy of 10 lm-eval tasks (arc_challenge, arc_easy, boolq, hellaswag, lambada_openai, mmlu, openbookqa, piqa, truthfulqa_mc1, winogrande); lm_eval --model hf with gguf_file0.633 → 0.60796.0%source ↗

69 more in quant.json

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json