UseToolAI
Kimi logo
Writing & Textweb development, full-codebase analysis, long-form research

Kimi Review 2026

Moonshot AI's flagship assistant, now on K3 — the largest open-source model ever (2.8T params) and #1 on front-end coding arenas. Its own launch note admits it still trails Claude Fable 5 and GPT-5.6 Sol overall.

★★★★4/5

What Makes Kimi Unique

The front-end development champion — the first open-source model to top a public arena against Claude and GPT — with unusually honest messaging about where it still trails.

What is Kimi?

Kimi K3 launched on July 18, 2026 with an unusual opening move: Moonshot's own announcement said, in its first sentence, that K3 still trails the strongest closed models, Claude Fable 5 and GPT-5.6 Sol. That honesty is the right frame for this card, because K3's story is more specific than "China's biggest model." It is the largest open-source model ever released at 2.8 trillion parameters, and it topped Frontend Code Arena at Elo 1,679, ahead of Fable 5 and GPT-5.6 Sol. It is the first time an open model has beaten closed frontier models on a public arena, with 6 of 7 web development sub-arenas won. The ranking is preliminary and the vote count is lower than the incumbents, but the result is real.

The price story needs the same honesty. API output rose to ¥100 per million tokens, roughly 3.5x the previous K2.6 rate, and the sticker price only looks reasonable if you hit the Mooncake cache, which in coding workflows reportedly exceeds 90%. At that rate the effective input cost drops near ¥3.8 per million tokens, DeepSeek territory. For non-coding use, expect to pay closer to list. Weights were released publicly on July 27, though actually running a 2.8 trillion parameter model locally is a hardware project of its own. If you build front-end work, K3 is currently the best tool for it. If you just want a general assistant, DeepSeek stays the better value. Pricing checked August 16, 2026.

Key Features

  • K3 (released July 18, 2026): 2.8 trillion parameter MoE model, 896 experts with 16 active per token
  • #1 on Frontend Code Arena at Elo 1,679 — the first open-source model to beat closed frontier models there (Claude Fable 5 at 1,631, GPT-5.6 Sol at 1,618); ranking is marked preliminary
  • 1M token context window — whole codebases or books in one session
  • Native vision understanding, not a vision module bolted onto a text model
  • Mooncake caching architecture: 90%+ cache hit rates reported in coding workflows
  • Weights released to the public on July 27, 2026

Pros & Cons

✓ Pros

  • +Front-end development champion: 6 of 7 webdev sub-arenas won, the first open-source win over closed models
  • +Refreshingly honest marketing: Moonshot's own launch note states K3 trails Claude Fable 5 and GPT-5.6 Sol overall
  • +Coding economics: at 90% cache hits, effective input cost lands near ¥3.8 per million tokens
  • +1M context is a real workflow change for long documents and full-repo analysis

✗ Cons

  • API prices rose roughly 3.5x over K2.6, with output now ¥100 per million tokens
  • Overall capability sits third: Artificial Analysis Index 57, behind Fable 5's 60 and Sol's 59 — the gap shows in general use
  • The cache advantage is a coding scenario; non-coding usage pays closer to the list price
  • Local deployment needs serious hardware — 2.8 trillion parameters does not run on a laptop, MoE or not

Who Is It Best For?

Front-end developers and people who work with very long documents. Best for: web development, full-codebase analysis, long-form research. Not for: general chat where DeepSeek offers better value, or work needing enterprise compliance.

Top Alternatives to Kimi

More in Writing & Text