HolyGrAIL

All tasks / Models

Kimi K3

Moonshot AI's open-weights Kimi MoE line; Kimi K3 has 2.8T total and 104B activated parameters and 1M context.

  • Moonshot AI
  • Open weights
  • Released Jul 2026

Official page

Where it ranks

  • S
    Chat & assistants
    Models · Local · 1,487 on LMArena Text Arena (overall, style control)
  • A
    Writing
    Models · Local · 1,461 on LMArena Text Arena: Creative Writing (style control)
  • –
    Documents & OCR
    Models · Local · no result on this task’s benchmarks yet
  • B
    Coding
    Models · Local · 1,655 on LMArena WebDev / Code Arena (overall)
  • A
    Image understanding
    Models · Local · 80.5% on Artificial Analysis MMMU-Pro (independent evaluation) (secondary benchmark)
  • C
    Vector graphics & logos
    Models · Local · 1,300 on Design Arena SVG
  • C
    Workflow automation
    Models · Local · 58.3% on AutomationBench-AA (Artificial Analysis)
  • A
    Slides & presentations
    Models · Local · 1,273 on Design Arena Python-PPTX Slides

Facts

Vendor
Moonshot AI
Family
Kimi
Released
Jul 2026
Weights
Open: download and run it yourself
License
Kimi K3 License
Parameters
2800B (104B active)
GPU memory to run locally
~1600 GB (4-bit ≈ 2.8 T × 0.57 GB ≈ 1.6 TB, estimate (multi-node datacenter).)
API price
$3 in / $15 out per 1M tokens
Access
Free (open weights)
Last verified
18 Oct 2026

Price from openrouter.ai, 18 Oct 2026.

Benchmark results

BenchmarkResultEntry on the leaderboardAs ofSource
Artificial Analysis Intelligence Index44Kimi K3 (Max)
Raw value from the page's embedded model data; the site displays it rounded to an integer.
18 Oct 2026artificialanalysis.ai
Artificial Analysis MMMU-Pro (independent evaluation)80.5%Kimi K3 (Max)
Raw fraction 0.8052 from the model data embedded in https://artificialanalysis.ai/leaderboards/models, converted to percent; rank 25 of 124 entries with an MMMU-Pro result; no last-updated date shown.
19 Oct 2026artificialanalysis.ai
AutomationBench-AA (Artificial Analysis)58.3%Kimi K3 (Max)
AutomationBench-AA Score (share of each task's objectives completed, a task with any guardrail violation scoring zero); raw value from the page's embedded model data (displayed rounded); rank 22 of 32; Tasks Completed 24.81%, Objectives Completed 84.83%; no last-updated date shown, so as_of = retrieval date.
19 Oct 2026artificialanalysis.ai
Design Arena HTML Slides1,256Kimi K3
3,960 battles. Read from the site's own API (page renders client-side).
18 Oct 2026designarena.ai
Design Arena Python-PPTX Slides1,273Kimi K3
3,277 battles. Read from the site's own API (page renders client-side).
18 Oct 2026designarena.ai
Design Arena SVG1,300Kimi K3
Design Arena SVG, rank 12, 1070 battles; rows read from POST https://www.designarena.ai/api/leaderboard (arenaType 'models', category 'svg'); names from https://www.designarena.ai/api/registry
19 Oct 2026designarena.ai
EQ-Bench Creative Writing v32,082kimi-k3
Elo score from the leaderboard data embedded in https://eqbench.com/creative_writing.js (file last modified 2026-09-24); rank 5 of 140; rubric score 16.85.
24 Sep 2026eqbench.com
LMArena Text Arena (overall, style control)1,487Kimi K3 (Max)
From LMArena's official leaderboard-dataset (text_style_control, split latest, category overall, published 2026-10-15), which mirrors https://arena.ai/leaderboard/text (site blocks automated fetches).
15 Oct 2026huggingface.co
LMArena Text Arena: Creative Writing (style control)1,461Kimi K3 (Max)
From LMArena's official leaderboard-dataset (text_style_control, split latest, category creative_writing, published 2026-10-15), which mirrors https://arena.ai/leaderboard/text/creative-writing (site blocks automated fetches). Rank 27 of 414; 6646 votes; 95% CI 1453.1-1469.5.
15 Oct 2026huggingface.co
LMArena WebDev / Code Arena (overall)1,655Kimi K3 (Max)
From LMArena's official leaderboard-dataset (webdev = Code Arena, split latest, category overall, published 2026-10-18), which mirrors https://arena.ai/leaderboard/webdev. Rank 15; 19426 votes; 95% CI 1648.6-1660.6.
18 Oct 2026huggingface.co
MDPBench (multilingual document parsing)83.6%Kimi-K3
Overall (public split, page-level mean of text, formula and table scores) from the board's leaderboard.json (updated 2026-10-04).
4 Oct 2026huggingface.co
SWE-Bench Pro V2 (public set, full)97.7%Kimi-K3 (mini-swe-agent) max
SWE-Bench Pro V2 Full (642 tasks); row added 2026-10-07; 95% CI +/-0.9; harness in the variant name; read from the leaderboard page's embedded data.
19 Oct 2026labs.scale.com

Where you can use it

Earlier versions

  • Kimi K2.6 · —
  • Kimi K2.5 · 27 Jan 2026

Compare

Alternatives to Kimi K3

Sources