← Top 10

Kimi K3

Moonshot

The strongest open-weights model in the world.

Best open

Ranked #17 of 402 on LMArena’s overall board (style control) · retrieved 2026-09-19

The numbers

Arena Score (LMArena, style control)1484.8 (1479.5–1490, 20,987 votes)
Coding board score1538.4 (#7, 5,507 votes)
Intelligence Index (Artificial Analysis)43.8
Median output speed~37 tokens/sec
Price per 1M tokens in/out (Artificial Analysis)$3 / $15
Context window~1M
LicenseKimi K3 License
Released2026-07-16

Open weights — you can run this model yourself.

Exact source entries: LMArena “kimi-k3-max” · Artificial Analysis “Kimi K3 (max)”, retrieved 2026-09-19.

Try or learn more

Honest read

Best at

  • Arena Score 1485 — ahead of every other non-proprietary model on the overall board
  • #7 on the coding board (1538) — open weights that genuinely code
  • Intelligence Index 43.8, above several closed frontier models; 1,048,576-token context

Not the pick for

  • Measured output speed 37 tokens/second — the slowest card we publish
  • Pricier than most open rivals: $3 / $15 per million tokens
  • Custom 'Kimi K3 license' — check terms before commercial self-hosting

Choose it if you are

Self-hosters, fine-tuners and anyone who needs the weights, not just the API.

Rankings come from the sources above; this plain-language read is our editorial summary of those numbers.

Compare with the leaders

Best image Text-to-image, by blind human votes GPT-Image 2.5 (Sunburst) · Arena Score 1421 Best video Text-to-video, by blind human votes Gemini Omni 1.1 Flash · Arena Score 1515 Best code Coding, by blind human votes Claude Fable 5 · Arena Score 1552 Best reasoning Highest Intelligence Index we track Claude Fable 5.1 · Index 53.4 Best value Intelligence per dollar GLM 5.3 Flash · 176 pts / $1M