Kimi K3 ships its weights: 2.8T parameters, MXFP4, and a 1M-token context

Moonshot AI on Monday published the full weights of Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model that's now the largest open-weight release in history. The drop lands squarely in the middle of an unresolved Washington debate about whether Chinese open models should be restricted at all, and it didn't wait for that debate to finish.

The architecture activates 16 of 896 experts per token, roughly 1.8% of the pool at inference, with a 1-million-token context window and, per Forbes, a 6.3x decoding speedup at million-token contexts attributed to the model's KDA attention scheme. Moonshot claims a 2.5x scaling-efficiency gain over K2. On the Arena blind Frontend Code eval, K3 ranked first at 1,679 points. On BrowseComp it scored 91.2, ahead of GPT-5.6 Sol at 90.4 and Fable 5 at 88.0.

The economics are the more interesting story. Per-benchmark rollout cost runs $4.65 for K3 against $13.41 for Fable 5, roughly 2.8x more solved tasks per dollar. Moonshot's API is priced at $0.30 per million cache-hit input tokens, $3 cache-miss, and $15 output. That uncached input number is five times K2's launch price. Cached-prompt gains, Moonshot claims, reach 10x versus Anthropic and OpenAI. The company is training the market to prepay context.

Markets read it immediately. Z.ai's Hong Kong shares shed as much as 30% during launch week; MiniMax fell 16%. Moonshot's daily revenue has grown at least sixfold since release, and the company is reportedly raising at a $50 billion valuation. Founder Yang Zhilin now controls what may be the most-downloaded frontier-class model on earth.

The political frame is unfinished. White House OSTP Director Michael Kratsios has to reconcile a growing constituency, Nvidia plus 24 other signatories to an open-weights letter, with a Trump administration instinct to restrict Chinese AI on principle. K3 will be hosted on AWS Bedrock, Azure Foundry, and Google Vertex AI regardless of what Washington concludes, because the weights are already out. Bank of America's Alex Liu and Bloomberg Intelligence's Mandeep Singh and William Tong have all flagged the same structural read: the open-weight center of gravity has moved east, and the U.S. hyperscalers are now distributors of a Chinese base model.

The historical rhyme is the 2017 release of Google's Transformer paper, which handed the field its architecture and let competitors compound on it for eight years. K3 does the same trick in reverse, and Moonshot posted the receipts before regulators finished reading the memo.

Sources

• https://www.bloomberg.com/news/articles/2026-07-27/china-s-moonshot-to-release-breakthrough-ai-model-for-download
• https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-ai-releases-weights-for-kimi-k3-firing-a-shot-across-the-bow-of-openai-and-anthropic-open-weight-model-performs-almost-as-well-as-frontier-models-while-being-2-3x-easier-to-run
• https://qz.com/moonshot-ai-kimi-k3-open-weights-download-072726
• https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
• https://www.forbes.com/sites/geruiwang/2026/07/27/why-kimi-k3-signals-a-convergence-toward-open-weight-models/