Moonshot Releases Kimi K2.6, a Trillion-Parameter Open Model

Moonshot AI released Kimi K2.6 on April 20, putting another trillion-parameter open-weight model into circulation and pushing the agentic-coding fight a step further into open territory.

The model is a mixture-of-experts design with one trillion total parameters and 32 billion active per token, paired with a 256K context window. Weights and code are published on Hugging Face under a Modified MIT license, which keeps the model usable for commercial work without the bolted-on restrictions some recent open releases now carry.

What sets K2.6 apart in practice is its agentic envelope. Moonshot says the model can run continuous coding sessions for up to thirteen hours, and that its agent cluster can support 300 sub-agents working in parallel across roughly four thousand coordinated steps. That is a long-horizon profile aimed at production engineering pipelines rather than chat demos.

The benchmark numbers

K2.6 scores 83.2 on BrowseComp, 80.2 on SWE-Bench Verified, and 66.7 on Terminal-Bench 2.0, according to MarkTechPost's coverage. Those numbers put the model within negotiating distance of GPT-5.4 and Claude Opus 4.6 on the agentic-coding workloads that enterprise buyers actually care about.

Distribution is unusually wide for a release week. The model is live on Kimi.com, the Kimi App, the public API and the Kimi Code CLI, and Cloudflare added it to Workers AI on the same day under the firm's serverless inference tier.

Where this fits

The release lands in a busy month for the open ecosystem. Mistral pushed Medium 3.5 with its own agent mode in early May, and Google's TurboQuant work on KV-cache compression is being absorbed into the open ecosystem at the same time. The shape of 2026's open-weight stack is filling in quickly, and Kimi K2.6 is a sizeable piece of it.