Back to timeline

Fri, July 1704:19Model/APIModel releasesAPI & pricingChinese modelsModel releases guide

Moonshot AI Releases 2.8T-Parameter Kimi K3, Pledges Open Source in 2026

Decision Brief

What changedMoonshot AI launched Kimi K3 with 2.8 trillion parameters, available via web and API, with open-source weights promised by July 27, 2027.
Why it mattersK3's 2.8T parameters and higher pricing ($3/M input, $15/M output) mean developers must assess API costs, but its Elo in long-context knowledge tasks trails only Claude Fable 5, and it tops frontend code benchmarks, making it attractive for teams needing high-quality code generation.
Who should careTeams building on model APIs
Affected stackClaudeOpenAIKimi
Builder actionWorth evaluating: can Claude replace or complement your current model
Source confidenceMedium · Reliable media or first-hand reporting

Moonshot AI released Kimi K3 on July 16, 2026, boasting 2.8 trillion parameters (approximately 3 trillion), making it the company's most powerful model to date. It is currently available via website and API, with open-source weights promised for July 27, 2027. Pricing is $3/million tokens input and $15/million tokens output, comparable to Anthropic Claude Sonnet series, making it the most expensive domestic model, significantly higher than its predecessor K2.6 ($0.95/$4). According to Artificial Analysis, K3 achieves an Elo of 1547 in long-context knowledge work evaluation, up 732 points from K2.6, second only to Claude Fable 5. Its per-task cost of $0.94 is similar to GPT-5.6 Sol ($1.04) and half of Claude Opus 4.8 ($1.80), but higher than open-source competitors. K3 also topped the frontend code arena on Arena.ai, surpassing Claude Fable 5. Internal benchmarks show K3 generally outperforms Claude Opus 4.8 max and GPT-5.5 high, but lags behind Claude Fable 5 and GPT-5.6 Sol. In a test generating SVG with K3, a single request consumed 95 input tokens and 16,658 output tokens (including 13,241 reasoning tokens), costing $0.25. The model currently only offers maximum reasoning effort, and system prompts may add ~85 tokens. For developers needing API-based long-context reasoning or frontend code generation, K3's high cost versus strong performance requires careful evaluation. Those seeking open-source models must wait until 2027.

Summary basis: official / RSS sourceCompiled from the source scope noted above; the original remains authoritative.

Sources

  • Simon Willison:Blog

    Hands-on notes on LLM tools, local models, and practical AI engineering.

  • Simon Willison:Blog

Related intel

留言

登入后即可留言,和其他 builder 交换实测心得。

还没有留言,抢头香。