Open Source · JULY 18, 2026
Moonshot's Kimi K3 lands at 2.8T parameters, sits behind only Fable 5 and GPT-5.6 Sol
Beijing's Moonshot AI unveiled Kimi K3 on July 16 — a 2.8-trillion-parameter sparse MoE with a 1M-token context window, novel Kimi Delta Attention, and full weights due July 27. It is the largest open-weight model ever released.
Beijing-based Moonshot AI unveiled Kimi K3 on July 16, a 2.8-trillion-parameter sparse mixture-of-experts model that sits behind only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on the leaderboards it targets, and the largest open-weight model anyone has ever released. Full weights are scheduled for July 27.
The number that matters isn't 2.8 trillion. It's 16. K3 activates just 16 of its 896 experts per token, roughly 1.8% of the pool, which is how Moonshot claims a 2.5x scaling-efficiency gain over Kimi K2. The architectural work carries names now: Kimi Delta Attention, a hybrid linear scheme, and something Moonshot calls Attention Residuals. There's a 1-million-token context window, MXFP4 weights and MXFP8 activations from quantization-aware training, and a homegrown compiler, MiniTriton, that Moonshot benchmarks against Triton on Nvidia H200s and the cut-down L20 that Nvidia is permitted to sell into China under U.S. export controls.
That last detail is the point of the exercise. Alex Liu, a Bank of America analyst quoted via a CNBC note, put it plainly: "Large-scale pre-training plus architectural work can still deliver step-change gains for flagship Chinese models despite compute constraints." K3 is the empirical form of that claim.
The competitive damage was immediate and legible in equity prices. Zhipu, listed in Hong Kong, fell 27.7% before Friday's close. MiniMax dropped 16.5%. K3 clears the previous Chinese open-ecosystem ceiling, set by Meituan's LongCat-2.0 and DeepSeek's V4-Pro at 1.6 trillion parameters each, by more than a full trillion. On Arena.ai's Frontend Code evaluation, K3 posted 1,679 points, ahead of Claude Fable 5 on that specific benchmark. Vals AI and Artificial Analysis coverage tracks a similar shape: K3 is within a few points of frontier US models on the tasks it was tuned for, and behind on others.
The economics tell a second story. K2 launched at $0.60 per million input tokens. K3's API asks $3 per million cache-miss input tokens (five times as much), $0.30 with a cache hit, and $15 per million output. Running the weights locally is quoted at hundreds of thousands of dollars of hardware. Open-weight in name; enterprise-priced in practice. Moonshot is meanwhile raising $2 billion at a roughly $30 billion valuation, per Bloomberg, which is the number that funds the next training run.
Ryan Fedasiuk of the American Enterprise Institute is among the analysts framing this as evidence that the export-control regime slows Chinese frontier work without stopping it. The gap that mattered a year ago was measured in orders of magnitude. It's now measured in benchmark points.
Sources
- https://www.bloomberg.com/news/articles/2026-07-17/china-s-powerful-new-moonshot-ai-model-closes-gap-with-us-rivals
- https://finance.yahoo.com/technology/ai/articles/chinas-moonshot-unveils-worlds-largest-020622030.html
- https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3
- https://techcrunch.com/2026/07/16/moonshots-upcoming-kimi-3-is-expected-to-close-the-gap-with-anthropics-opus-4-8/
- https://platform.kimi.ai/docs/guide/kimi-k3-quickstart