Kimi K2.6

What it is

Kimi K2.6 is an open-weight model released by Moonshot AI in April 2026 ^[1]. It is a mixture-of-experts model with 1 trillion total parameters and 32 billion active per token, using 384 experts with 8 selected and 1 shared, and it has a 262,144-token context window ^[2]. It accepts text and image input. Video input is experimental and available only through Moonshot's official API ^[2].

Origins

Moonshot released it under a modified MIT license ^[2].

The model is focused on long-horizon coding, meaning software work that runs over many steps, in languages such as Rust, Go and Python. It is also aimed at turning prompts and visual inputs into interactive interfaces ^[2]. The 262,144-token context window is the amount of text, code and images the model can consider at once, so a large codebase or a long document can stay in view during a single task.

Main features

The agent system was expanded in this version. Agent Swarm scales to 300 sub-agents and 4,000 coordinated steps. The model also supports persistent background agents for scheduled tasks and keeps its thinking across turns of a conversation ^[2]. A persistent background agent keeps running on its own rather than waiting for a user message, which is what makes scheduled tasks possible. Keeping the thinking across turns means the model's earlier reasoning remains available when a conversation continues, rather than being discarded at the end of each turn.

Moonshot reported these scores: SWE-Bench Verified 80.2, SWE-Bench Pro 58.6, Terminal-Bench 2.0 66.7, LiveCodeBench v6 89.6 and HLE-Full with tools 54.0 ^[2]. Other reported results are 83.2 on BrowseComp (86.3 with Agent Swarm), 96.4 on AIME 2026, 90.5 on GPQA-Diamond and 73.1 on OSWorld-Verified ^[2]. OSWorld-Verified tests how well a model can operate a computer desktop, and GPQA-Diamond is a set of hard graduate-level science questions.

Its place in development

Kimi K2.6 was followed in June 2026 by Kimi K2.7 Code, a coding-specialized version with a 256K context and multi-step tool invocation ^[1]. Moonshot has also reported a long-horizon evaluation Elo for the next generation, Kimi K3, that is 732 points above K2.6 ^[3].