Moonshot AI published the full weights for Kimi K3 on Monday, July 27, eleven days after the model's initial unveil and one day after it took the top slot on Arena.ai's Frontend Code leaderboard at 1,679 Elo, ahead of Anthropic's Claude Fable 5. At 2.8 trillion total parameters, K3 is the largest open-weight model ever released, and the first Chinese release that competitive benchmarking treats as a peer of GPT-5.6 Sol and Claude Opus 4.8 rather than a discount alternative to them.

The architecture is the story. K3 is a mixture-of-experts model with 896 experts, of which only 16 are activated per token, or roughly 1.8% of the pool. Active parameters come in at 104.2 billion, a shape that gives Moonshot dense-model quality at sparse-model serving cost. The company claims 2.5x scaling efficiency over Kimi K2, and pairs the MoE routing with Kimi Delta Attention (a hybrid linear-attention scheme with a fixed-size state handler) and a feature it calls Attention Residuals. Context window is one million tokens. Weights ship as MXFP4 with MXFP8 activations, and Moonshot recommends supernodes of 64 or more accelerators for serving.

Pricing is where the geopolitics gets legible. Moonshot is charging $3 per million cache-miss input tokens, $0.30 on cache hits, and $15 for output. That's frontier performance at pricing that would embarrass any US lab's margin structure.

The kernel work is worth noting too. Moonshot benchmarks its MiniTriton compiler against Triton, and Tom's Hardware reports the optimization work was done on Nvidia H200s at an undisclosed location, with the served model running on the export-compliant L20. The L20 is a cut-down Ada part designed specifically for the Chinese market. That K3 runs on it at all is the practical rebuttal to the export-control thesis.

Washington noticed. The Nasdaq dropped about 1% on the July 17 announcement, the same day Xi Jinping addressed the World AI Conference in Shanghai. David Sacks, co-chair of the President's Council of Advisors on Science and Technology, said the U.S. is "tying itself in knots" while Chinese labs ship. Dean Ball, OpenAI's head of strategic futures, warned that the release would "create large amounts of regulatory risk around the use of open-weight Chinese models." Anthropic, which in February accused Moonshot of training on 3.4 million Claude exchanges, has said nothing publicly about the new release.

The subtext of Ball's line is the interesting part. When a US lab's head of strategic futures frames an open-weight competitor as regulatory risk rather than technical competition, the argument has already moved off the merits. That's the shape of every incumbent response to a credible entrant, from the browser wars forward. K3 has made the argument portable.

Sources

https://www.bloomberg.com/news/articles/2026-07-27/china-s-moonshot-to-release-breakthrough-ai-model-for-download
https://www.bloomberg.com/news/articles/2026-07-17/china-s-powerful-new-moonshot-ai-model-closes-gap-with-us-rivals
https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
https://techcrunch.com/2026/07/18/kimi-threat-or-menace/
https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3