whatmodel.sh / MiniMax-M3

MiniMax-M3

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

MiniMax-M3 is an open-weights language model from MiniMax, released June 2026. Its published size is 428B parameters (23B active per token) and a 1M-token context window. Licence: minimax-community.

70.2±1.0#79 of 719 LLMsfrom 16 benchmark results across 8 sources

$ how it compares

The LLMs ranked either side of MiniMax-M3, out of 828 with a score, then more from MiniMax.

#modelscorewhereRAMreleased
82Qwen3.5 397B-A17B Alibaba71±0api $1.09/M +local~65GB2026-02
83Ornith 1.5 35B-A3B DeepReinforce71±2local~22GB2026-08
84MiniMax-M3 MiniMax70±1api $0.48/M +local~137GB2026-06
85Grok 4.3 SpaceXAI (xAI)70±1api $1.56/M2026-04
86Gemini 2.5 Pro Preview (06-05) Google70±2comparison only2025-06
more from MiniMax
120MiniMax-M2.7 MiniMax68±1api $0.53/M +local~75GB2026-03
150MiniMax-M2.5 MiniMax66±1api $0.53/M +local~75GB2026-02
154MiniMax-M2.1 MiniMax66±1local~79GB2025-12

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #57 of 300 · 35% of its score · 5 tests

Beats Gemini 3.1 Pro (Preview), GPT-5.2, Gemini 3.6 Flash here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Arena Agent (October 2026 aggregate IPS)-0.0687 IPS ✎▁▁▁▄▂▃▇█▂▂▁▃ beats 8/49avg -0.312 IPS▇▇▇▇▇▇▇ 18%sharp
MCP-Atlas (public set)74.2% ✎█··█·██·████ beats 3/7hard · avg 29%▇▇ 5.8%at its level
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)19.4% ✎█▄·█·▄·██▄·▄ beats 2/11very hard · avg 4.7%▇▇ 5.8%typical
Legal Research Bench (agentic, all-pass)29.8% ✎▄█··▄▄▄██·▄█ beats 7/12very hard · avg 5.8%▇ 2.8%typical
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)53.0% ✎▂·▁▁··▇▇█▃·▁ beats 24/25hard · avg 29%▇ 2.6%blunt · at its level
coding▁▁▁▃▆██▇▆▄▃▁ #56 of 458 * · 21% of its score · 2 tests

Beats GPT-5.4, DeepSeek V4 Flash Vision (Exp), Gemini 3.1 Pro (Preview) here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1482.0 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 67/119avg 981▇▇▇▇▇▇ 16%sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)19.4% ✎█▄·█·▄·██▄·▄ beats 2/11very hard · avg 4.7%▇▇ 5.8%typical
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #66 of 393 * · 8.8% of its score · 1 test

Beats GPT-5.2, Qwen3.8 27B, LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails Nemotron Cascade 2 30B-A3B, Gemma 4 31B, Gemini 2.5 Pro here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1440.1 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 298/379avg 1268▇▇▇▇ 8.8%blunt
math▁▂▃▅▆▅▄█▄▁▁▁ #139 of 398 * · 4.0% of its score · 1 test

Beats Claude Sonnet 4.6 here, though they rank higher overall. Trails Olmo 3 32B Think, Olmo 3.1 32B Think, gpt-oss-20b here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
OTIS mock AIME 2024–202571.1% ✎▃▃▄▃▃▃▇▇█▃▂▅ beats 105/193very hard · avg 8.5%▇▇ 4.0%sharp · noisy
vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #59 of 147 · 9.1% of its score · 3 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-5 here, though they rank higher overall. Trails Gemma 4 26B-A4B, Qwen3.6 Flash, Qwen3.5 35B-A3B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena vision (style controlled)1236.6 ✎▁▁▂▂▂▂▄▅▅▆██ beats 92/145avg 1068▇▇ 6.1%blunt
Arena document (style controlled)1433.0 ✎▁▁▃▃▄▃▇█▅▇▄▅ beats 5/37avg 1335▇ 2.9%blunt
PureDocBench document parsing (Avg3 over Clean/Digital/Real)50.7% ✎▃·▁·▂▂▅█·▁▂▂ beats 1/22medium · avg 62%▇ 0.2%blunt · at its level
web search▂▁▂▆▄█▃▃▁▁·▁ #43 of 80 * · 2.8% of its score · 1 test

Beats Gemini 3.1 Pro (Preview), Qwen3.7 Max, GPT-5.2 here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Legal Research Bench (agentic, all-pass)29.8% ✎▄█··▄▄▄██·▄█ beats 7/12very hard · avg 5.8%▇ 2.8%typical
other tests5 not in a task · 27% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond90.9% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 180/208hard · avg 40%▇▇▇▇ 11%sharp
Mystery Game Puzzles (Epoch)7.0% ✎█··▂▃▇▄█▂▃▁▁ beats 13/74very hard · avg 9.4%▇▇ 5.8%sharp · noisy
Chess Puzzles (Epoch)14.0% ✎█▁▁▁▁▃▁▄▂▂▁▁ beats 67/135very hard · avg 5.9%▇▇ 4.9%sharp · noisy
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)31.3% ✎▁▁▃▂▂▄▃▃▄▄▄█ beats 13/38very hard · avg 14%▇ 3.0%at its level
Legal Research Bench (single-shot, no tools, all-pass)2.9% ✎▃▃█▃·▅···█▃▃ beats 3/12very hard · avg 0.8%▇ 3.0%too hard for it

$ where to run it

Through an API

api · Deep Infra · $0.49/M · $0.28 in / $1.1 out per 1M tokens · 1M context · deepinfra.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 MiniMax-M3 · needs ~257.3 GB · from huggingface.co/unsloth
    chat in the terminal
    llama cli -hf unsloth/MiniMax-M3-GGUF:UD-Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf unsloth/MiniMax-M3-GGUF:UD-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 MiniMax-M3 · needs ~257.3 GB · from huggingface.co/unsloth

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/unsloth/MiniMax-M3-GGUF@UD-Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

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 MiniMax-M3 · needs ~241.5 GB, Mac only · from huggingface.co/mlx-community
    chat in the terminal
    mlx_lm.chat --model mlx-community/MiniMax-M3-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model mlx-community/MiniMax-M3-4bit

Official weights and documentation: huggingface.co/MiniMaxAI ↗

$ how the score was worked out · LMArena leaderboard dataset 51%, Epoch AI — internal evaluatio… 25%, Legal Research Bench (Vals AI… 6% and 5 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
LMArena leaderboard dataset5▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 51%CC BY 4.0
Epoch AI — internal evaluations4▇▇▇▇▇▇▇▇ 25%CC BY 4.0
zai-org/GLM-5.2 model card1▇▇ 6%mit
GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions (arXiv:2609.21562v2)1▇▇ 6%CC BY 4.0
Miðeind Icelandic LLM Leaderboard1▇ 3%CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)1▇ 3%CC BY 4.0
How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)1▇ 0%CC BY 4.0
prior · pull toward the middle▇ 0%
weighted average= 70.2

$ 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