whatmodel.sh / Kimi K3

Kimi K3

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

Kimi K3 is an open-weights language model from Moonshot AI, released July 2026. Its published size is 2800B parameters (104B active per token) and a 1M-token context window. Licence: kimi-k3-license.

77.2±0.6#25 of 719 LLMsfrom 28 benchmark results across 12 sources

$ how it compares

The LLMs ranked either side of Kimi K3, out of 828 with a score, then more from Moonshot AI.

#modelscorewhereRAMreleased
23Gemini 3.8 Flash Google77±1api $1.50/M2026-09
24Muse Spark 1.3 Meta77±1api $2.00/M2026-09
25Kimi K3 Moonshot AI77±1api $6.00/M +local~884GB2026-07
26GPT-5.2 Pro OpenAI77±1api $57.75/M2025-12
27MiMo-V2.6-Flash (309B-A15B) Xiaomi77±1api $0.17/M +local~136GB2026-09
more from Moonshot AI
61Kimi K2.6 Moonshot AI73±1api $1.71/M +local~317GB2026-04
72Kimi K2.7 Code Moonshot AI72±1api $1.71/M +local~359GB2026-06
92Kimi K2.5 Moonshot AI70±1api $1.20/M +local~160GB2026-01

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #23 of 300 · 17% of its score · 6 tests

Beats GPT-5.5, GPT-5.6 Terra, GPT-5.4 here, though they rank higher overall. Trails Grok 4.7, Qwen3.8 Max (0902), GLM-5.3 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena Agent (October 2026 aggregate IPS)0.0418 IPS ✎▁▁▁▄▂▃▇█▂▂▁▃ beats 37/49avg -0.312 IPS▇▇▇ 6.3%sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)36.1% ✎█▄·█·▄·██▄·▄ beats 6/11very hard · avg 4.7%▇ 2.6%at its level
τ³-bench banking_knowledge (v1.0.1 grading)37.1% ✎▃▃▂▃·▆▃▂█▅▅▂ beats 14/25very hard · avg 4.1%▇ 2.5%at its level
E2E-SWE (build a working codebase from scratch, 186 tasks) pass@142.3% ✎▃··▅··▃█▃▅▃▅ beats 4/12very hard · avg 3.1%▇ 2.4%sharp · at its level
Thinkingbox-bench pass@1 (507 stateful business-workflow agent tasks, N=20)57.4% ✎▅··▃·▃·█▃▅██ beats 10/15very hard · avg 4.9%▇ 1.5%at its level
Video2World V2WScore (coding agents reconstruct interactive simulated worlds from videos, OpenCode scaffold)33.2% ✎█··▄█·▄····▄ beats 5/6very hard · avg 6.6%▇ 1.5%at its level
coding▁▁▁▃▆██▇▆▄▃▁ #13 of 458 · 13% of its score · 4 tests

Beats GPT-5.6 Sol, Claude Fable 5, GPT-5.5 here, though they rank higher overall. Trails Qwen3.8 Max (0902) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1657.8 ✎▁▂▄▃▅█▄▇▆▄▁▁ beats 109/119avg 981▇▇ 5.6%sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)36.1% ✎█▄·█·▄·██▄·▄ beats 6/11very hard · avg 4.7%▇ 2.6%at its level
E2E-SWE (build a working codebase from scratch, 186 tasks) pass@142.3% ✎▃··▅··▃█▃▅▃▅ beats 4/12very hard · avg 3.1%▇ 2.4%sharp · at its level
Luau Leetcode fully solved (Torpedo Software)77.9% ✎▁▁▁▁·▁▂▃▅▃█▇ beats 28/49hard · avg 18%▇ 1.9%typical
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #12 of 393 * · 3.1% of its score · 1 test

Beats GPT-6 Astra, GPT-6.1 Sol, Claude Sonnet 5.5 here, though they rank higher overall. Trails Claude Opus 4.6, Muse Spark 1.1, Claude Opus 4.7 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1488.4 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 367/379avg 1268▇ 3.1%blunt
math▁▂▃▅▆▅▄█▄▃▁▁ #23 of 398 · 52% of its score · 10 tests

Beats DeepSeek V4.1 Flash, Gemini 3.8 Flash here, though they rank higher overall. Trails GPT-5.6 Luna, GPT-6 Luna, GPT-5.2 Pro here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
FrontierMath tier 4 v239.0% ✎▁▂▁▂█▅▂▂▁▁▁▁ beats 42/63very hard · avg 0.1%▇▇▇▇▇▇ 14%sharp · at its level
FrontierMath tiers 1–3 v272.2% ✎▂··▁▁▃▄▄█▅▃▃ beats 61/82very hard · avg 1.2%▇▇▇▇▇ 13%sharp · at its level
MathArena ArXivMath June 2026 (48 problems)71.5% ✎▂···▄·██▆▆▄▂ beats 12/19very hard · avg 4.6%▇▇▇ 7.5%sharp · at its level
MathArena ArXivMath April 2026 (40 problems)54.2% ✎▁·····▅██▄·▁ beats 10/20very hard · avg 7.2%▇▇▇ 6.8%sharp · at its level
MathArena BrokenArXiv April 2026 (61 problems)59.4% ✎▂▂▆█▃▃▃▂▂··▂ beats 15/18very hard · avg 0.7%▇ 3.1%sharp · at its level · noisy
MathArena ArXivMath May 2026 (40 problems)61.7% ✎▁··▁▄█▄▁·▃·▁ beats 11/17very hard · avg 10%▇ 2.4%at its level
MathArena BrokenArXiv June 2026 (54 problems)51.9% ✎▄··█▂▂▂▆▄··▂ beats 11/14very hard · avg 1.1%▇ 2.1%sharp · at its level · noisy
MathArena ArXivMath August 2026, with tools (57 problems)38.6% ✎█▄·▄·······█ beats 0/5very hard · avg 4.8%▇ 1.6%sharp · at its level
2 more tests
testresultvs the others in its rundifficultycountswhy
MathArena BrokenArXiv August 2026, with tools (56 problems)61.9% ✎▄▄·····▄·▄·█ beats 2/5very hard · avg 5.5%▇ 0.7%at its level · noisy
OTIS mock AIME 2024–202597.2% ✎▃▃▄▃▃▃▇▇█▃▃▅ beats 169/193very hard · avg 8.5%▇ 0.4%sharp · maxed out · noisy
vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #24 of 147 * · 1.2% of its score · 1 test

Beats Claude Sonnet 5.5, MiMo-V2.6-Pro, GPT-5.6 Terra here, though they rank higher overall. Trails Gemini 3.6 Flash, Gemini 3.5 Flash, GLM-5.3-Flash here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
CharXiv (reasoning questions) lab84.8% ✎▂▅▃▅▅▆▆▅▆█▆▂ beats 93/100medium · avg 39%▇ 1.2%small run
other tests8 not in a task · 20% of its score
testresultvs the others in its rundifficultycountswhy
Furniture Assembly (Epoch)34.2% ✎█▆▆▆▆▂▅▂▂▂·▅ beats 12/30very hard · avg 1.2%▇▇ 5.6%sharp · at its level
Mystery Game Puzzles (Epoch)26.0% ✎█··▂▃▇▄█▂▃▁▁ beats 44/74very hard · avg 9.4%▇▇ 5.0%sharp · at its level · noisy
Chess Puzzles (Epoch)39.0% ✎█▁▁▁▁▂▁▄▂▃▁▁ beats 116/135very hard · avg 5.9%▇ 3.0%sharp · at its level · noisy
GPQA Diamond93.1% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 193/208hard · avg 40%▇ 1.9%sharp
SimpleQA Verified50.6% ✎▁▁▃▂▃▃▆▆█▃▂▅ beats 55/78very hard · avg 8.0%▇ 1.3%at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)40.7% ✎▁▁▃▂▂▄▁▃▄▄▄█ beats 14/38very hard · avg 14%▇ 1.1%at its level
AutomationBench public set (Zapier README)46.7% ✎▄······▄█··▄ beats 3/4very hard · avg 10%▇ 1.1%at its level
RobloxQA v2.0 multiple choice (Torpedo Software)96.5% ✎▁▁▃▁█▄▃▃▆▂▂▂ beats 34/49easy · avg 78%▇ 0.6%maxed out

$ where to run it

Through an API

api · $6.00/M · $3 in / $15 out per 1M tokens · 1M context · platform.moonshot.ai ↗ · 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 Kimi K3 · needs ~1680.5 GB · from huggingface.co/unsloth
    chat in the terminal
    llama cli -hf unsloth/Kimi-K3-GGUF:UD-Q4_K_XL
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf unsloth/Kimi-K3-GGUF:UD-Q4_K_XL

LM Studio

version

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

    download it (macOS / Windows / Linux) ↗

  2. Run Kimi K3 · needs ~1680.5 GB · from huggingface.co/unsloth

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/unsloth/Kimi-K3-GGUF@UD-Q4_K_XL -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Official weights and documentation: huggingface.co/moonshotai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 44%, MathArena ArXivMath / BrokenA… 24%, LMArena leaderboard dataset 15% and 9 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
Epoch AI — internal evaluations8▇▇▇▇▇▇▇▇▇▇▇▇▇ 44%CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets7▇▇▇▇▇▇▇ 24%CC BY-SA 4.0
LMArena leaderboard dataset3▇▇▇▇ 15%CC BY 4.0
GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions (arXiv:2609.21562v2)1▇ 3%CC BY 4.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode2▇ 3%MIT
τ²/τ³-bench leaderboard: Sierra's own evaluation runs1▇ 2%MIT
E2E-SWE: Benchmarking LLMs on Building Working Codebases from Scratch (arXiv:2609.38335v1)1▇ 2%CC BY 4.0
One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)1▇ 2%CC BY 4.0
Video2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos (arXiv:2610.04432v2)1▇ 1%CC BY-SA 4.0
Moonshot AI vendor-reported benchmark scores for its own Kimi models (Hugging Face model cards) · lab-reported1▇ 1%Factual vendor-reported score (not licensed content)
Miðeind Icelandic LLM Leaderboard1▇ 1%CC BY 4.0
Zapier AutomationBench README: public-set pass rates1▇ 1%MIT
prior · pull toward the middle▇ 0%
weighted average= 77.2

$ 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