What Was Released
<cite index="20-14,20-15">Moonshot AI released free, public download weights for Kimi K3 on July 26, 2026, at roughly 7:30 PM EDT — a day ahead of the July 27 target the company had previously communicated — with confirmed specifications of 2.8 trillion parameters and a 1,048,576-token (1M) context window.</cite> <cite index="23-4">The July 27 date marked the planned publication of the model's full weights on Hugging Face, the step that allows organizations with sufficient infrastructure to download, inspect, fine-tune, and self-host the model under a Modified MIT license.</cite>
<cite index="20-17">Cloud platforms moved fast: Together AI and Modal both announced day-0 hosted access, timed to the drop.</cite> The weights are available in Moonshot's official Hugging Face repository.
Architecture
<cite index="10-3">Kimi K3 is the world's first open-weight 2.8-trillion-parameter artificial intelligence (AI) model, introducing Kimi Delta Attention (KDA), Stable LatentMoE, and a 1M-token context window.</cite> <cite index="11-13,11-14">With the Stable LatentMoE framework, the model efficiently activates 16 out of 896 experts. Together with improvements in training methodology and data recipes, these structural advances give Kimi K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively.</cite> <cite index="27-18">The MXFP4 weights total roughly 1.4 TB.</cite>
Benchmark Performance
<cite index="12-5">Moonshot said K3 still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance, but it outperformed every other model in the company's evaluation suite, including Claude Opus 4.8 and GPT 5.5, across coding and agentic benchmarks.</cite> <cite index="12-9">Arena ranked K3 first in its Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing.</cite> <cite index="15-10,15-11">Artificial Analysis gives K3 a score of 57 on its Intelligence Index and ranks it fourth among 189 models; it sits behind Claude Fable 5 and two GPT-5.6 Sol reasoning settings, then ahead of Claude Opus 4.8, GPT-5.5 at xhigh, Claude Sonnet 5, and GLM-5.2.</cite>
<cite index="3-14">The key fact is that either the open-to-closed or American-to-Chinese model performance gap has been reduced from the debated 6–9 months to something shorter, say 3–5 months.</cite>
Company Background and Strategy
<cite index="4-4">Moonshot AI was founded in 2023 by Yang Zhilin, a Tsinghua University graduate who previously conducted research at Google and Meta.</cite> <cite index="2-2">The company is backed by Alibaba and Tencent.</cite> <cite index="4-8,4-9">The release of DeepSeek's low-cost R1 model in January 2025 disrupted the entire Chinese AI landscape, and Moonshot AI was among the hardest hit — Kimi, which had ranked third in monthly active users in China, slid to seventh.</cite>
<cite index="2-10">Moonshot's shift toward open-weight releases began with Kimi K2 in July 2025 and accelerated with K2.5 in January 2026, a strategy the company has used to rebuild its position in China's AI market.</cite> <cite index="1-8">Annual recurring revenue (ARR) reached approximately $200 million in April 2026 and $300 million in June, a 50% increase in two months driven by enterprise subscriptions and API (Application Programming Interface) services, per Bloomberg reporting.</cite> <cite index="1-10">Moonshot raised $2 billion at a $20 billion valuation in a Meituan-led round in May 2026, per Bloomberg's May 2026 funding report.</cite>
API Pricing and Access
<cite index="12-3">API pricing is $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens.</cite> <cite index="2-4">The API is compatible with the OpenAI software development kit (SDK), which lowers the integration barrier for developers already building on OpenAI or Anthropic tools.</cite>
Geopolitical Context
<cite index="6-4">K3's performance underscores a recurring pattern: three years of escalating restrictions on graphics processing units (GPUs) and lithography equipment have not prevented Chinese labs from reaching or nearing the frontier.</cite> <cite index="21-4,21-5,21-6">On July 22, White House Office of Science and Technology Policy (OSTP) Director Mike Kratsios accused Moonshot of running an internal distillation platform against Claude Fable 5 using restricted Nvidia chips acquired via Thailand; Treasury Secretary Scott Bessent warned sanctions and Entity List designations are possible; skeptics note K3 testing reportedly predated Fable 5's release.</cite>
<cite index="2-11">Moonshot's approach reflects a broader pattern among Chinese AI developers, including DeepSeek, Alibaba, Tencent, and Baidu, all of which have released open-weight models over the past year.</cite>