MDL-9379EST.2026 · IDX.738
Language modelIn production

Kimi K2.6

Moonshot AI · China · 2026

Moonshot AI's trillion-parameter open model: a multimodal agent that other companies rebrand and resell as their own national product.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Kimi K2.6, published on 20 April 2026, is the release that made Moonshot AI's open line the default foundation for other vendors' products — Japan's Sakana AI, for instance, ships it retuned as Sakana Namazu. It is a mixture-of-experts model with one trillion total parameters and 32 billion active per token, 384 experts of which eight are selected per token, 61 layers, a 256K context window and native vision through a 400M-parameter MoonViT encoder, so images enter the model directly rather than through a separate captioning step. Moonshot aims the release squarely at agent work: the card claims long-horizon coding that holds up across Rust, Go and Python, the ability to turn prompts and screenshots into working interfaces, and an agent swarm that scales to 300 sub-agents running 4,000 coordinated steps in a single autonomous run. Reported results include 54.0 on HLE-Full with tools and 83.2 on BrowseComp, in both cases above the frontier closed models Moonshot chose for comparison. Weights are published on Hugging Face under a modified MIT licence, which is what makes the model attractive to companies that want to sell an adapted version without owning a training run: at roughly $0.95 per million input tokens through resellers it costs a fraction of the closed models it is benchmarked against. The catalogue also holds the later Kimi K3.

#open weights#mixture of experts#agentic#multimodal#modified MIT licence
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review