AIBRIEF
Moonshot publishes the trained model for free and prices its API at roughly one third of Anthropic's Claude Fable 5; it ranks third on the Artificial Analysis intelligence index.
A Beijing startup released a 2.8-trillion-parameter AI model on Monday and published the trained weights for anyone to download. Moonshot AI's Kimi K3 uses a Mixture-of-Experts architecture, which only activates part of the model per query, and reads up to a million tokens of text in one pass. The GitHub repository and Hugging Face model card include the weights and a technical report.
Kimi K3 ranks third on the Artificial Analysis independent intelligence index, behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. Its API pricing runs at roughly one-third of Claude Fable 5 for comparable coding work. Moonshot separately claims a 250% lift in compute efficiency on constrained hardware; that figure comes from the company's own statement and has not been independently verified.
The same week, Alibaba's Qwen team previewed Qwen3.8, a roughly 2.4-trillion-parameter system, with the open-weights release forthcoming.
The honest caveat: a 2.8-trillion-parameter model is not something most readers will run on a laptop. Production use still means hitting the API or a hosted cluster. What changes is price and access: open-weight has now reached the closed-source frontier on cost, not just on capability.