Nathan Lambert's independent analysis frames K3 as a frontier-class open-weights escalation: Moonshot AI has either reduced the open-to-closed or US-to-Chinese model performance gap from a debated 6–9 months down to 3–5 months, and K3 is the closest an open model has been to the frontier since DeepSeek R1. Lambert argu The piece carries strategic caveats for developers and providers: much of the conclusion rests on Moonshot honoring the promised July 27, 2026 open-weights release; if K3 stays closed, the open-closed gap rebounds. Lambert also flags that any adversarial distillation contribution from closed frontier US models is at mo

Omid Saffari's July 17, 2026 review frames Kimi K3 as Moonshot's flagship for coding, agents and end-to-end knowledge work, accessible through Kimi.com, Kimi Work, Kimi Code and the kimi-k3 API, and explicitly cautions that K3 wins only selected routes against Claude Fable 5 and GPT-5.6 Sol. The piece restates K3's API Saffari emphasizes that K3 is a mixture-of-experts model rather than a fully dense 2.8T-parameter one, and that at launch only the "max" effort variant is exposed while weights and the technical report remain unpublished. The takeaway decision rule is to trial K3 when cost constrains a long-context agent route and to k

Independent hands-on coverage of Moonshot AI's Kimi K3 launch (announced Jul 16, 2026) describes it as a 2.8-trillion-parameter model, marketed as the first "open 3T-class" release, with an open-weight release promised by July 27, 2026. Reported pricing is $3 per million input tokens and $15 per million output tokens — According to Artificial Analysis figures cited in the post, Kimi K3 reaches an overall Elo of 1547 on a private long-horizon knowledge-work evaluation (+732 over K2.6, behind only Claude Fable 5), with cost per task of $0.94 (vs GPT-5.6 Sol's $1.04 and roughly half of Opus 4.8's $1.80) and a 21% reduction in output-tok

The official Kimi Code "What's New" changelog documents three recent CLI releases: v0.37.0 (Aug 18, 2026), v0.38.0 (Aug 20, 2026), and v0.39.0 (Aug 27, 2026). v0.39.0 introduces experimental Remote Control for local web sessions (env flag KIMI CODE EXPERIMENTAL REMOTE CONTROL=1), an experimental Tower multi-agent orche v0.38.0 adds a WaitFor tool so the agent can block on a background task within a single turn, plus 13 new data sources in the official Kimi Datasource plugin (Chinese government data from NDA/NBS, standards GB/HB/DB/TT, eight international datasets — WHO, FAO, UNSD, ECB, Eurostat, UNICEF, OECD, FRED — and Xinhua Financ