whatmodel.sh / Gemma 4 31B

Gemma 4 31B

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

Gemma 4 31B is an open-weights language model from Google, released April 2026. Its published size is 30.7B parameters and a 262k-token context window. Licence: apache-2.0.

66.9±0.6#116 of 719 LLMsfrom 74 benchmark results across 26 sources

$ how it compares

The LLMs ranked either side of Gemma 4 31B, out of 828 with a score, then more from Google.

#modelscorewhereRAMreleased
131Qwen3.7 Flash Alibaba67±1api $0.055/M2026-07
132Step 3.7 Flash (198B-A11B) StepFun67±1api $0.44/M +local~104GB2026-05
133Gemma 4 31B Google67±1api $0.19/M +local~21GB2026-04
134Gemini 3.1 Flash-Lite Google67±1api $0.56/M2026-03
135Qwen3-Coder-Next (80B-A3B) Alibaba67±2api $0.60/M +local~27GB2026-02
more from Google
9Gemini 4 Argon (limited preview) Google82±1limited preview2026-09
23Gemini 3.8 Flash Google77±1api $1.50/M2026-09
32Gemini 3.7 Flash Google76±1api $1.50/M2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #107 of 300 · 35% of its score · 17 tests

Beats Grok 4.3, o3-pro, Qwen3.6 Plus here, though they rank higher overall. Trails SmolLM3 3B, Nova 2 Pro (Preview), Mistral Medium 3.5 (128B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Pro (split unspecified) +133.0% ✎▄···▄···▄▄██ beats 2/7very hard · avg 6.5%▇▇ 5.0%sharp · at its level · several runs
Terminal-Bench Hard lab +136.0% ✎██··█·█·█··█ beats 5/5very hard · avg 4.7%▇▇ 4.9%sharp · at its level · several runs
SWE-bench Verified +160.6% ✎▄▄·····▄▄▄▄█ beats 4/7hard · avg 17%▇▇ 4.8%sharp · at its level · several runs
Terminal-Bench 2.130.3% ✎▄······█▄▄█▄ beats 3/7very hard · avg 7.1%▇ 3.4%at its level
ContractBench success rate (33 observation-contract tool tasks)37.4% ✎▆··▆▄▂▄▂▆▆█▄ beats 9/23very hard · avg 14%▇ 2.6%sharp · at its level
SWE-bench Pro public set, Qwen-refined +135.7% ✎▄······▄█·▄▄ beats 0/5hard · avg 18%▇ 2.3%at its level · several runs
τ²-bench Telecom lab69.3% ✎▄····█··▄·▄▄ beats 5/5very hard · avg 14%▇ 2.0%sharp · at its level · noisy
SWE-bench Multilingual49.3% ✎▄·▄·····█▄█▄ beats 3/7hard · avg 18%▇ 1.6%at its level
9 more tests
testresultvs the others in its rundifficultycountswhy
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)23.3% ✎▃····█··▃█·▃ beats 4/8very hard · avg 3.4%▇ 1.5%typical
Tau2 – airline lab75.0% ✎▃·▃··▃·····█ beats 4/5medium · avg 42%▇ 1.4%typical
τ²-bench aggregate lab76.9% ✎▄·▄··▄····█▄ beats 5/5hard · avg 32%▇ 1.2%at its level
OSWorld-Verified58.5% ✎█▄·▂▆▂▄▂▆▄▆█ beats 9/25hard · avg 26%▇ 1.2%at its level · small run
MCP-Atlas (public set)54.2% ✎▂·▁▂▁·▁▃█▆▆▂ beats 7/37hard · avg 28%▇ 0.9%at its level · small run
DeepSearchQA61.7% ✎█·······█··█ beats 0/2medium · avg 40%▇ 0.8%at its level · small run
Humanity's Last Exam with tools lab26.5% ✎▂▂▂▃▄▂▃▆█▇▂▆ beats 7/54very hard · avg 11%▇ 0.7%at its level · small run
Tau2 – retail lab86.4% ✎▄····▄··▄·▄█ beats 5/5medium · avg 40%▇ 0.4%sharp · noisy
T1-Bench multi-domain conversational agent pass^341.1% ✎▂▂··▂·▂·▂▂█▄ beats 11/11hard · avg 25%▇ 0.0%blunt · at its level · noisy
coding▁▁▁▃▆██▇▆▄▃▁ #108 of 458 · 24% of its score · 9 tests

Beats Grok 4.3, Qwen3.5 Plus, DeepSeek V3.2 here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4, Laguna XS 2.1 (33B-A3B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Pro (split unspecified) +133.0% ✎▄···▄···▄▄██ beats 2/7very hard · avg 6.5%▇▇ 5.0%sharp · at its level · several runs
SWE-bench Verified +160.6% ✎▄▄·····▄▄▄▄█ beats 4/7hard · avg 17%▇▇ 4.8%sharp · at its level · several runs
Arena WebDev (Code Arena)1364.7 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 35/119avg 981▇▇ 4.8%sharp
Luau Leetcode fully solved (Torpedo Software)55.3% ✎▁▁▁▁·▁▂▃▅▃█▇ beats 12/49hard · avg 18%▇ 2.4%at its level
SWE-bench Pro public set, Qwen-refined +135.7% ✎▄······▄█·▄▄ beats 0/5hard · avg 18%▇ 2.3%at its level · several runs
SWE-bench Multilingual49.3% ✎▄·▄·····█▄█▄ beats 3/7hard · avg 18%▇ 1.6%at its level
SciCode (level unspecified)33.2% ✎▄···▄▄█▄·▄·▄ beats 5/7very hard · avg 13%▇ 1.3%at its level
SciCode lab43.0% ✎█▄·······▄▄▄ beats 5/5hard · avg 21%▇ 1.3%at its level
1 more test
testresultvs the others in its rundifficultycountswhy
LiveCodeBench v6 (Feb–May 2025) lab80.0% ✎▄··▄·▄···▄·█ beats 5/5medium · avg 43%▇ 0.8%typical
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #52 of 393 * · 3.0% of its score · 2 tests

Beats MiMo-V2.6-Flash (309B-A15B), GPT-6 Luna, GPT-5.6 Luna here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1452.8 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 317/379avg 1268▇ 2.7%blunt
IFEval lab98.9% ✎▄··▄··█···▄▄ beats 5/5very easy · avg 94%▇ 0.3%maxed out
math▁▂▃▅▆▅▄█▄▁▁▁ #116 of 398 · 3.9% of its score · 3 tests

Beats Claude Sonnet 4.6, MiniMax-M3, MiMo-V2.5 (310B-A15B) here, though they rank higher overall. Trails Olmo 3.1 32B Think, Gemini 2.5 Flash, Qwen3-VL 32B Thinking here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
OTIS mock AIME 2024–202573.3% ✎▃▃▄▃▃▃▇▇█▃▂▅ beats 109/193very hard · avg 8.5%▇ 2.0%sharp · noisy
MATH-Vision lab85.6% ✎██·█·····███ beats 5/5medium · avg 53%▇ 1.0%typical
AIME 2026 lab89.2% ✎▄·▄▄····▄··█ beats 5/5medium · avg 41%▇ 0.9%sharp · noisy
vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #47 of 147 · 12% of its score · 11 tests

Beats MiMo-V2.6-Pro, DeepSeek V4 Flash Vision (Exp), GPT-5 here, though they rank higher overall. Trails Qwen3.6 Flash here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena vision (style controlled)1262.3 ✎▁▁▂▂▂▂▄▅▅▆██ beats 110/145avg 1068▇ 1.9%blunt
ChartArena chart parsing (English, mAP_high average)46.8% ✎▁·▁·▃▃▁▇▃█·▄ beats 18/22hard · avg 27%▇ 1.5%blunt · at its level
MMMU Pro lab76.9% ✎██·█····█·██ beats 5/5medium · avg 49%▇ 1.5%at its level
MATH-Vision lab85.6% ✎██·█·····███ beats 5/5medium · avg 53%▇ 1.0%typical
BEAR-Bench (bilingual RU/EN enterprise and academic document images, overall accuracy)23.5% ✎▃▃·▅·▃·█·█▅▅ beats 2/14hard · avg 24%▇ 1.0%at its level
CharXiv (reasoning questions)77.7% ✎▂▅▃▅▅▆▆▅▆█▆▂ beats 70/100medium · avg 42%▇ 1.0%at its level · small run
OmniDocBench 1.5 overall (higher is better)72.5% ✎▂··▃▂▂··▂▂█▅ beats 1/13medium · avg 54%▇ 0.9%small run
InfographicVQA lab92.0% ✎▄█·······█·▄ beats 5/5easy · avg 71%▇ 0.9%typical
3 more tests
testresultvs the others in its rundifficultycountswhy
Arena document (style controlled)1444.5 ✎▁▁▃▃▄▃▇█▅▇▄▅ beats 9/37avg 1335▇ 0.9%blunt
OmniDocBench 1.5 ↓ lab13.1% ✎█·█·····████ beats 5/5avg 31%▇ 0.6%blunt
ScreenSpot-Pro75.9% ✎▂▃▅▆██▅▂···▂ beats 21/24medium · avg 51%▇ 0.5%blunt · at its level · small run
web search▂▁▂▆▄█▃▃▁▁·▁ #66 of 80 · 3.1% of its score · 3 tests

Beats DeepSeek V4 Flash (Preview, Apr 2026), Nemotron 3 Ultra 550B-A55B, GLM-4.7 here, though they rank higher overall. Trails GPT-4.1 nano, GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)23.3% ✎▃····█··▃█·▃ beats 4/8very hard · avg 3.4%▇ 1.5%typical
DeepSearchQA61.7% ✎█·······█··█ beats 0/2medium · avg 40%▇ 0.8%at its level · small run
Humanity's Last Exam with tools lab26.5% ✎▂▂▂▃▄▂▃▆█▇▂▆ beats 7/54very hard · avg 11%▇ 0.7%at its level · small run
other tests31 not in a task · 37% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond +175.8% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 116/208hard · avg 40%▇▇▇ 6.6%sharp · at its level · several runs
MedXPertQA MM lab61.3% ✎▂···▂·▃··▆▂█ beats 9/13very hard · avg 30%▇ 2.8%sharp · at its level
NL2Repo +115.5% ✎▄·····█·▄▄·▄ beats 0/5very hard · avg 8.2%▇ 2.6%at its level · several runs
EQ-Bench 3 (emotional intelligence, judge Elo)1272.9 ✎▁·▂▁▄▅▅▃█▄▁▁ beats 41/62avg 1044▇ 1.7%typical
MRCR v2 8-needle, 128k lab66.4% ✎▄▄·▄···█···▄ beats 5/5hard · avg 22%▇ 1.6%at its level
BRIDGE clinical text, zero-shot average (87 tasks)46.7% ✎▁▁▁▁▁▂▂▃▄▆█▃ beats 96/97hard · avg 32%▇ 1.5%blunt · at its level
MMLU-Pro lab85.2% ✎▄··█···▄··▄▄ beats 5/5medium · avg 61%▇ 1.3%typical
MMMLU lab88.4% ✎██··█···█·██ beats 5/5medium · avg 69%▇ 1.3%typical
23 more tests
testresultvs the others in its rundifficultycountswhy
AraGen-v3 3C3H score (Inception, Claude Sonnet 4 judge)72.7% ✎▁▁·▁▂▂▄█▄▅▄▄ beats 57/66hard · avg 31%▇ 1.3%at its level
Big Bench Extra Hard micro avg lab74.4% ✎█··▄···▄·▄·▄ beats 5/5hard · avg 28%▇ 1.2%at its level
IFBench lab76.0% ✎▄▄·▄······▄█ beats 5/5medium · avg 40%▇ 1.2%at its level
QwenClawBench (Qwen internal)41.7% ✎▄·▄····▄···█ beats 1/4hard · avg 22%▇ 1.1%at its level
WildClawBench37.6% ✎▂·▄▄▆▃▄█▆▂▂▂ beats 8/20very hard · avg 14%▇ 1.1%at its level · small run
Chess Puzzles (Epoch)5.0% ✎█▁▁▁▁▂▁▄▂▂▁▁ beats 18/135very hard · avg 5.9%▇ 1.0%sharp · noisy
CPTUB average (SpeakLeash, 0-5 judge scale)86.2% ✎▁▂▂▂▃█▃▆▅▄▂▃ beats 91/95easy · avg 66%▇ 1.0%typical
RobloxQA v2.0 multiple choice (Torpedo Software)92.6% ✎▁▁▃▁█▄▃▃▆▂▂▂ beats 22/49easy · avg 78%▇ 1.0%typical
SimpleQA Verified10.4% ✎▁▃▃▂▃▃▆▆█▃▂▅ beats 2/78very hard · avg 8.0%▇ 0.9%at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)20.7% ✎▁▁▃▂▃▄▁▃▄▄▄█ beats 8/38very hard · avg 14%▇ 0.9%at its level
Open Persian LLM Leaderboard v2 average (21 tasks)78.5% ✎▂▄▂▃█▄▃▆▃▅▂▃ beats 81/81medium · avg 58%▇ 0.9%blunt
EuroEval Life in the UK (English knowledge MC), validation split, 5-shot, accuracy95.2% ✎▂·▂▂·▂·▃▃·█▃ beats 14/14easy · avg 80%▇ 0.8%maxed out
EuroEval HellaSwag (English), validation split, 5-shot generative MC, accuracy89.1% ✎▂··▂▂·▂▆▆█·▄ beats 15/15medium · avg 63%▇ 0.8%typical
Gaia236.4% ✎█·····█····█ beats 0/2hard · avg 16%▇ 0.7%at its level · small run
SkillsBench (with skills)32.4% ✎█·········██ beats 0/2hard · avg 16%▇ 0.7%at its level · small run
HLE lab19.5% ✎█···█······█ beats 2/2very hard · avg 3.9%▇ 0.6%small run
SkillsBench 78-task subset (OpenCode, avg of 5)23.6% ✎▄·····▄···▄█ beats 2/4very hard · avg 7.1%▇ 0.6%typical
DiagnosticIQ industrial maintenance MCQ (macro accuracy, 4 options)65.1% ✎▁··▃▁▁▁▅█▅▁▅ beats 17/24medium · avg 52%▇ 0.5%blunt · at its level
WMT25 Terminology Translation (term success rate)92.1% ✎▂·▂··▂·▄▂·█▆ beats 12/12medium · avg 64%▇ 0.5%noisy
NTCIR SciClaimEval Subtask 1 pair-accuracy (dev, single model)69.9% ✎▃▃···▃·▃▃█·▃ beats 5/8medium · avg 45%▇ 0.4%blunt · at its level
WMT24++79.2% ✎█·▃·▅▅▃···▃▃ beats 8/10easy · avg 71%▇ 0.2%blunt
WMT26 General MT (GEMBA-ESA, LLM judge)79.5% ✎▄·▄▄▄·▄██▄█▄ beats 9/12easy · avg 65%▇ 0.2%blunt
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)92.6% ✎▁▂▁▂▅█▂▃▄▃▁▁ beats 66/98very easy · avg 90%▇ 0.0%blunt

$ where to run it

Through an API

api · DeepInfra · $0.19/M · $0.13 in / $0.38 out per 1M tokens · 262k context · deepinfra.com ↗ · price sourceUSD per 1M text tokens, standard uncached input and output. Product: Serverless text generation. API model: google/gemma-4-31B-it. Settings: service tier=standard, batch=no, cache=uncached input. Additional/excluded charges: Tools, cache storage, priority service and other modalities excluded.. Currency: USD.

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 Gemma 4 31B · needs ~18.7 GB · from huggingface.co/lmstudio-community
    chat in the terminal
    llama cli -hf lmstudio-community/gemma-4-31B-it-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf lmstudio-community/gemma-4-31B-it-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 Gemma 4 31B · needs ~18.7 GB · from huggingface.co/lmstudio-community

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/lmstudio-community/gemma-4-31B-it-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 Gemma 4 31B · needs ~18.92 GB · from ollama.com
    chat in the terminal
    ollama run gemma4:31b

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 Gemma 4 31B · needs ~18.4 GB, Mac only · from huggingface.co/lmstudio-community
    chat in the terminal
    mlx_lm.chat --model lmstudio-community/gemma-4-31B-it-MLX-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model lmstudio-community/gemma-4-31B-it-MLX-4bit

Official weights and documentation: huggingface.co/google ↗

$ how the score was worked out · Gemma Team — Gemma 4 Technica… 27%, Kwaipilot/KAT-Coder-V2.5-Dev… 12%, meta-models/Muse-Glimmer-30B… 12% and 23 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
Gemma Team — Gemma 4 Technical Report · lab-reported20▇▇▇▇▇▇▇▇ 27%CC BY 4.0
Kwaipilot/KAT-Coder-V2.5-Dev model card5▇▇▇▇ 12%Apache-2.0
meta-models/Muse-Glimmer-30B model card (speed and benchmark tables)11▇▇▇▇ 12%Apache-2.0
LMArena leaderboard dataset4▇▇▇ 10%CC BY 4.0
Epoch AI — internal evaluations4▇▇ 8%CC BY 4.0
Qwen/Qwen3.6-35B-A3B model card4▇ 4%apache-2.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode2▇ 3%MIT
Qwen/Qwen3.6-27B model card2▇ 3%apache-2.0
ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)1▇ 3%CC BY 4.0
google/gemma-4-31B-it model card · lab-reported2▇ 2%Apache-2.0
EQ-Bench 3 canonical leaderboard results1▇ 2%MIT
NVIDIA ModelOpt quantized-checkpoint cards: baseline (unquantized) measurements of third-party models1▇ 2%Apache-2.0 (Gemma-4-31B-IT-NVFP4 repo) / MIT (DeepSeek-V4-Pro-0813-NVFP4 repo)
EuroEval leaderboards (results collection), English generative leaderboard2▇ 2%MIT
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)1▇ 2%MIT
ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)1▇ 1%CC BY 4.0
K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts (arXiv:2606.02404v1)1▇ 1%CC BY 4.0
Inception Arabic/Hindi leaderboards (AraGen-v3, HindiGen-v1, Arabic IFEval)1▇ 1%Apache-2.0
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)1▇ 1%MIT
BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)1▇ 1%CC BY 4.0
Miðeind Icelandic LLM Leaderboard1▇ 1%CC BY 4.0
Open Persian LLM Leaderboard v2 (Part DP AI and AUT NLP Lab)1▇ 1%Apache-2.0
Cohere — North Small Translate (arXiv:2609.13916v1)3▇ 1%CC BY 4.0
DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules (arXiv:2605.08614v1)1▇ 1%CC BY 4.0
SciTrue: Reliable Scientific Claim Validation with Frontier and Open Language Models at the NTCIR SciClaimEval Task — authors' own single-model dev evaluation1▇ 0%CC BY 4.0
T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains (arXiv:2606.11070v1)1▇ 0%CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)1▇ 0%Apache-2.0
prior · pull toward the middle▇ 0%
weighted average= 66.9
$ published speed measurements · 1

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

AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GBQ4_0 (Google QAT GGUF per guide)11.38 tok/sllama.cpp 86a9c79f8 (build_number 10066 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence

Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 8 (2026-07-18, measured-local): Official current Google QAT GGUF direct r3 scout. Text vision and native tool-call smokes passed separately. Practical workstation-state compatibility row with a known low-load CPU-only background workload; not a strict-clean speed headline or broad quality evaluation.

source ↗
$ quantization evidence · 35 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
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)AIME 2025 0-shot pass@1, mean of 8 seeds (without thinking)0.658 → 0.683103.8%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)LiveCodeBench v6 0-shot pass@1 (without thinking)0.714 → 0.735102.9%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)GPQA Diamond 0-shot pass@1 (without thinking)0.774 → 0.779100.7%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)IFEval 0-shot prompt-level strict (without thinking)0.907 → 0.912100.6%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)GPQA Diamond 0-shot pass@1 (with thinking)0.864 → 0.869100.6%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)AIME 2025 0-shot pass@1, mean of 8 seeds (with thinking)0.871 → 0.875100.5%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)MATH-500 0-shot pass@1 (with thinking)0.879 → 0.881100.2%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)MMLU-Pro 0-shot custom-extract (without thinking)0.854 → 0.854100.0%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)MMLU-Pro 0-shot custom-extract (with thinking)0.871 → 0.87100.0%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)IFEval 0-shot prompt-level strict (with thinking)0.946 → 0.94299.6%source ↗
FP8 block-wise (data-free; llm-compressor model_free_ptq FP8_BLOCK)MATH-500 0-shot pass@1 (without thinking)0.894 → 0.88799.2%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)LiveCodeBench v6 0-shot pass@1 (with thinking)0.743 → 0.792106.7%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)AIME 2025 0-shot pass@1, mean of 8 seeds (with thinking)0.871 → 0.9103.4%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)MATH-500 0-shot pass@1 (with thinking)0.879 → 0.895101.8%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)MATH-500 0-shot pass@1 (without thinking)0.894 → 0.903101.0%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)GPQA Diamond 0-shot pass@1 (without thinking)0.774 → 0.781100.9%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)AIME 2025 0-shot pass@1, mean of 8 seeds (without thinking)0.658 → 0.662100.6%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)IFEval 0-shot prompt-level strict (without thinking)0.907 → 0.911100.4%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)IFEval 0-shot prompt-level strict (with thinking)0.946 → 0.949100.3%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)MMLU-Pro 0-shot custom-extract (with thinking)0.871 → 0.871100.0%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)MMLU-Pro 0-shot custom-extract (without thinking)0.854 → 0.85399.9%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)GPQA Diamond 0-shot pass@1 (with thinking)0.864 → 0.8699.6%source ↗
FP8 dynamic (data-free; FP8 weights per-channel, activations dynamic per-token; llm-compressor model_free_ptq FP8_DYNAMIC)LiveCodeBench v6 0-shot pass@1 (without thinking)0.714 → 0.70798.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)LiveCodeBench v6 0-shot pass@1 (with thinking)0.743 → 0.779104.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)GPQA Diamond 0-shot pass@1 (with thinking)0.864 → 0.87100.8%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)IFEval 0-shot prompt-level strict (with thinking)0.946 → 0.9499.3%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)MMLU-Pro 0-shot custom-extract (with thinking)0.871 → 0.86499.2%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)AIME 2025 0-shot pass@1, mean of 8 seeds (with thinking)0.871 → 0.86399.0%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)IFEval 0-shot prompt-level strict (without thinking)0.907 → 0.89899.0%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)MATH-500 0-shot pass@1 (with thinking)0.879 → 0.8798.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)LiveCodeBench v6 0-shot pass@1 (without thinking)0.714 → 0.70798.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)MMLU-Pro 0-shot custom-extract (without thinking)0.854 → 0.84598.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)GPQA Diamond 0-shot pass@1 (without thinking)0.774 → 0.76698.9%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)AIME 2025 0-shot pass@1, mean of 8 seeds (without thinking)0.658 → 0.6598.7%source ↗
NVFP4 (FP4 weights/activations with per-group scaling; llm-compressor NVFP4)MATH-500 0-shot pass@1 (without thinking)0.894 → 0.85195.2%source ↗

$ 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