Kimi K2.6

How it is built

Kimi K2.6 has the shape of Kimi K2.5: 1T parameters, 32B active per token.¹¹1Moonshot AI, Kimi K2.6 model card, April 2026. Runs with thinking on, temperature 1.0, top-p 1.0 and a 262,144-token context. SWE-Bench Pro and Verified with an in-house framework adapted from SWE-agent, coding scores averaged over 10 runs; Terminal-Bench 2.0 with Terminus-2; BrowseComp with the agent swarm; HLE-Full with search, code-interpreter and web-browsing tools. https://huggingface.co/moonshotai/Kimi-K2.6 Moonshot says a K2.5 deployment can be reused as it is.¹¹

Sixty of the 61 layers carry experts: 384 routed and 1 shared in each, with 8 routed experts chosen per token.¹¹ An expert is 2,048 wide.¹¹

Attention is Multi-head Latent Attention (MLA), with 64 heads at a hidden size of 7,168.¹¹ The vocabulary has 160K tokens. Images enter through MoonViT, an encoder of 400M parameters.¹¹

What is new

The Kimi blog dates K2.6 to 20 April 2026.²²2Moonshot AI, Kimi blog index, entry dated 2026-04-20. https://www.kimi.ai/blog

Moonshot measured it against K2.5. SWE-Bench Pro rose from 50.7 to 58.6, averaged over 10 runs in an in-house framework adapted from SWE-agent. That framework gives the agent six tools, bash among them.¹¹ In the same framework SWE-Bench Verified went from 76.8 to 80.2.¹¹ The larger jump is on Terminal-Bench 2.0 with the Terminus-2 agent, from 50.8 to 66.7. That run kept thinking preserved across turns and used the default JSON parser.¹¹

The agent swarm scales to 300 sub-agents working through 4,000 coordinated steps.¹¹ With the swarm, BrowseComp reaches 86.3. With search and code tools, HLE-Full rose from 50.2 to 54.0.¹¹

Thinking can be switched off. The card calls that instant mode and recommends a temperature of 0.6 for it.¹¹

Running it

Modified MIT is the MIT licence with one condition. Above 100 million monthly active users, or US$20 million of monthly revenue, a product has to show "Kimi K2.6".³³3Moonshot AI, Modified MIT License of Kimi K2.6. https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/LICENSE

The weights are native INT4, quantized the way Kimi K2 Thinking was.¹¹ Moonshot suggests vLLM or SGLang, and offers a Kimi Vendor Verifier to check that a third-party deployment behaves correctly.¹¹ Video input is experimental and works only on Moonshot's own API.¹¹

kimi-k2.6 costs $0.95 per million input tokens, or $0.16 on a cache hit, and $4.00 per million output tokens. Its context is 262,144 tokens.⁴⁴4Kimi API Platform, chat model pricing, per million tokens, as listed on 29 September 2026. https://platform.kimi.ai/docs/pricing/chat

Kimi K2.7 Code is a coding model built on K2.6.⁵⁵5Moonshot AI, Kimi K2.7 Code model card: "a coding-focused agentic model built upon Kimi K2.6". https://huggingface.co/moonshotai/Kimi-K2.7-Code