Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Model That Cracks the Global Top Three

Moonshot AI, the Beijing lab backed by Alibaba, Tencent, Meituan and HSG, released Kimi K3 on July 16 — a 2.8-trillion-parameter sparse mixture-of-experts model the company calls the largest open-source model ever built. It went live on Kimi.com and the API immediately; the weights themselves are promised for July 27, making K3, for now, an open-weights model in waiting.

The architecture

K3 activates just 16 of 896 experts per query and carries a 1-million-token context window. Two internally developed components do the heavy lifting: Kimi Delta Attention, a hybrid linear-attention mechanism that Moonshot says delivers up to 6.3x faster decoding at million-token lengths, and Attention Residuals, a drop-in replacement for residual connections the lab credits with consistent scaling gains. Reasoning is always on: the API exposes only a maximum reasoning_effort setting, with no way to disable thinking.

The scoreboard

On GDPval-AA v2, a benchmark measuring real-world tasks across 44 occupations and 9 industries, K3 scored 1,687 — third overall, behind only Anthropic's Claude Fable 5 Max (1,815) and OpenAI's GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600). It also took second place on AA-Briefcase with 1,527. Fortune's verdict was blunt: the release pushes Chinese AI into what it called Fable-level territory.

Efficiency as strategy

Moonshot president Yutong Zhang framed the design as a product of constraint: "We knew we didn't have the luxury to simply scale up compute. That forced us to focus on fundamental research and efficiency." API pricing lands at $0.30 per million cache-hit input tokens, $3.00 cache-miss and $15.00 output — undercutting Western frontier rates by a wide margin.

Timing is everything

The release landed on the eve of the World Artificial Intelligence Conference in Shanghai, which opened July 17 with Xi Jinping delivering the keynote. Moonshot is meanwhile reported to be raising at a $31.5 billion valuation, and the company's claim to the "largest open-source model" title comes with an asterisk until the July 27 weights drop actually happens.