Moonshot unveiled Kimi K3 on July 16 and 17 with 2.8 trillion parameters, a 1-million-token context window, and a promise to release the full model weights by July 27 (Moonshot). Arena's live WebDev table ranks kimi-k3 first with a preliminary score of 1,679 from 1,757 votes (Arena).
You cannot self-host it yet.
Access today means the hosted API. The download that lets an operator fine-tune, inspect, quantize, or host K3 on their own hardware remains scheduled for July 27. The leaderboard result therefore describes a model that most teams can evaluate only through Moonshot's infrastructure.
| Window | What you can do |
|---|---|
| July 17–26 | Call Moonshot's endpoint. No weights, no fine-tuning, no air-gapped install. |
| July 27 | Full-weight release and technical report scheduled by Moonshot. |
As of July 20, Artificial Analysis places K3 fourth of 187 on its Intelligence Index, behind Claude Fable 5 and GPT-5.6 Sol (Artificial Analysis). Arena labels the WebDev result preliminary. A team choosing a model for agentic planning or math should treat it as one signal about one skill.
Until the weights arrive, evaluating K3 means using Moonshot's hosted infrastructure. Teams with data-residency rules or code that cannot leave their network must wait for the download before running a meaningful internal test.
Watch July 27 for the weights, license, technical report, and supported quantizations. Those details will determine whether a 2.8T release is usable outside the largest inference shops.
The Signal is the public edge of a private practice. Sherpa points the same intelligence engine at one owner's business — competitors, suppliers, regulators, watched daily, graded and sourced. Work with a Sherpa →