Moonshot AI opens the weights of Kimi K2.6
On 14 April 2026 Moonshot AI published the weights of Kimi K2.6 on Hugging Face - a native multimodal agentic mixture-of-experts model with 1 trillion parameters (32 billion active per token) - under a modified MIT licence.
Why it matters
A model at the trillion-parameter scale, able to process both images and text, became available to run and fine-tune independently under a permissive licence. An editorial assessment.
By the Hugging Face model card, Kimi K2.6 is a mixture-of-experts decoder with 384 experts (8 selected per token, 1 shared), 61 layers, a 256K-token context, and a MoonViT vision encoder (400 million parameters) - so the model accepts both images and text. The repository moonshotai/Kimi-K2.6 on Hugging Face was created on 14 April 2026 (by the API timestamp); the licence is the company's own 'Modified MIT License'. Its own results table (the README as of the repository's creation) compares the model against GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro Preview and the earlier Kimi K2.5. What the record does not claim. The company's tech blog (kimi.ai/blog/kimi-k2-6.html), linked from the model card, carries no date and has since been extended with a comparison against Kimi K3, a later model released in July 2026; no figure from that blog was used in this record. The page names no hardware requirement for running the model independently.