Moonshot's Kimi K3 lands at 2.8T parameters, second only to Fable 5 on long-horizon agent tasks

The Beijing lab is calling it the first open 3T-class model. Full weights land July 27.

Moonshot AI announced Kimi K3 on Thursday, a 2.8-trillion-parameter model that VentureBeat is calling the largest open-source release to date, with full weights scheduled to drop July 27. On Artificial Analysis’s private AA-Briefcase evaluation for long-horizon knowledge work, it posts an Elo of 1,527, trailing only Anthropic’s Claude Fable 5 Max at 1,587 and clearing GPT-5.6 Sol Max’s 1,495.

The frame is what matters. A Beijing lab is now the second-best long-horizon agent on a private US benchmark, and it’s giving the weights away.

Simon Willison’s write-up pegs K3 at 1,547 on the public Artificial Analysis figure, a jump of 732 points from Kimi K2.6. On GDPval-AA v2, which spans 44 occupations, K3 scores 1,687. Architecturally, VentureBeat cites a 1-million-token context window, native vision, always-on “thinking mode,” and two Moonshot-internal components called Kimi Delta Attention and Attention Residuals. The API is compatible with the OpenAI SDK, which is the tell that Moonshot understands where the developer mindshare actually lives.

Pricing is the other lever. Fortune has K3 at $15 per million output tokens, against $50 for Fable, $4.40 for z.ai’s GLM-5.2, and $0.87 for DeepSeek V4. Willison, citing Artificial Analysis, puts cost-per-task at $0.94 for K3 versus $1.80 for Opus 4.8. Cursor’s Composer 2 already routes to Kimi. So does DoorDash, whose CTO Andy Fang said in early July that his team hands “lower-level work to Kimi K2.6.”

Moonshot describes K3 in its release: “Operating with minimal human oversight, it can sustain long engineering sessions, navigate massive repositories, and orchestrate terminal tools.” TechCrunch reports Moonshot is raising at a $31.5 billion valuation, up from $20 billion in May.

The 2023 DeepSeek-R1 moment reframed cost curves in a week. K3 is the sequel argument: that the frontier isn’t just cheap from Beijing, it’s open.

Moonshot's Kimi K3 lands at 2.8T parameters, second only to Fable 5 on long-horizon agent tasks

Sources