Moonshot announces Kimi K3, a 2.8-trillion-parameter open-weight model with reasoning always on

Moonshot announced K3 on 2026-07-16 as a 2.8T-parameter model priced at $3 / $15 per million tokens, with an open-weight release promised "by July 27, 2026" and, per Willison, only one reasoning…

Date
16 July 2026
Who
Moonshot AI
Confidence
Medium (launch details via Willison; model details from the Hugging Face card; w
Deep dive
Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them

Tier: Supporting · Significance: 3/5 · Org(s): Moonshot AI · Confidence: Medium (launch details via Willison; model details from the Hugging Face card; weights date unconfirmed) Moonshot announced K3 on 2026-07-16 as a 2.8T-parameter model priced at $3 / $15 per million tokens, with an open-weight release promised "by July 27, 2026" and, per Willison, only one reasoning effort, max, at launch (Simon Willison). The Hugging Face model card gives 104B active parameters (896 experts, 16 per token), a 1,048,576-token context, text, image and video understanding, and thinking that is always on with effort levels low, high and max (max the default); it is released under a custom Kimi K3 License under which Model-as-a-Service operators above US$20 million of revenue in any 12 months must negotiate a separate commercial agreement (model card, license), so "open weights" carries that condition. A secondary report dates the weights' posting 2026-07-26, which I did not confirm.

Artificial Analysis, quoted by Willison, placed K3 second only to Claude Fable 5 on its private Elo with self-reported scores mostly above Claude Opus 4.8 and GPT-5.5 (high) and below Fable 5 and GPT-5.6 Sol. In Das's IMO harness it needed multiple repair rounds and 17.4 hours to reach 42/42 (B08-48). Its predecessors are k1.5 and K2 Thinking; Anthropic's 2026 distillation reports name Moonshot (B08-14). For the wider picture see B20.

Read it in the deep dive