Moonshot AI Wraps Up Kimi K3 Rollout With Full Model Weight Release
Moonshot AI has finalized the rollout of its flagship Kimi K3 model by publicly releasing the complete set of open weights, marking the full availability of what the company positions as the first open-weight system in the three-trillion-parameter class.
The weights became downloadable on July 27, 2026, eleven days after the initial hosted service debut on July 16, fulfilling the timeline the Beijing-based firm announced at launch.
Kimi K3 is a sparse mixture-of-experts architecture containing 2.8 trillion total parameters, of which roughly 104 billion activate for any given token.
It draws from a pool of 896 experts and routes each token to 16 of them, delivering high capacity while keeping inference costs closer to those of far smaller dense models.
The system supports a one-million-token context window and processes text, images, and video natively through a vision encoder trained from scratch.
Architectural innovations include Kimi Delta Attention, a hybrid linear-attention design that occupies the majority of the model’s 93 layers, paired with Attention Residuals that selectively propagate information across depth.
Together with a stable latent mixture-of-experts framework, these changes yield an estimated 2.5-fold gain in scaling efficiency relative to the earlier Kimi K2 generation.
The model is engineered for extended, multi-step workloads.
It can maintain coherent engineering sessions across large code repositories, orchestrate terminal tools, and tackle tasks ranging from GPU kernel optimization and compiler work to vision-guided game development, CAD, and even chip design.
On the knowledge-work side it generates research reports complete with interactive visualizations, dashboards, and multimedia outputs.
Independent evaluations place it near the frontier: it ranks just behind leading proprietary systems such as Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on aggregate intelligence indexes while outperforming many peers on coding, agentic, and front-end web-development benchmarks.
Access began on the company’s own platforms—Kimi.com, the mobile apps, the Kimi Work desktop client, the Kimi Code terminal tool, and the Moonshot API—with pricing set at approximately $0.30 per million cached input tokens, $3 for uncached input, and $15 for output.
Demand quickly strained capacity, prompting a temporary pause on new subscriptions shortly after the initial release.
The open-weight checkpoint, quantized and sharded for practical distribution, now resides on Hugging Face and GitHub under the Kimi K3 License, which permits broad commercial and research use with attribution requirements for very large deployments.
Supporting infrastructure components and a detailed technical report were released concurrently, accelerating community efforts to integrate the model into inference engines such as vLLM and SGLang.
By making a near-frontier-scale system freely downloadable, Moonshot continues its pattern of pushing the upper limits of open-model size and capability.
The release lowers barriers for researchers, enterprises, and developers who wish to fine-tune, self-host, or build specialized agents without sole reliance on closed APIs, while underscoring the progress of Chinese AI labs in closing the gap with proprietary Western systems.
