Kimi K2.7 Code

How it is built

Moonshot built Kimi K2.7 Code on Kimi K2.6 and kept its architecture, down to the deployment method.¹¹1Moonshot AI, Kimi K2.7 Code model card, June 2026. Benchmarks with thinking on in Kimi Code CLI, temperature 1.0, top-p 0.95 and a 262,144-token context; GPT-5.5 in Codex at xhigh; MCP Atlas with the official configuration, averaged over 3 runs. https://huggingface.co/moonshotai/Kimi-K2.7-Code The size is unchanged: 1T parameters, 32B of them active for each token.¹¹

One of the 61 layers is dense, and every other layer has 384 routed experts plus 1 shared one. The router picks 8 per token.¹¹ Attention is MLA with 64 heads and a hidden size of 7,168. A MoonViT encoder of 400M parameters reads images.¹¹

What is new

The Kimi Code changelog dates the release to 12 June 2026. The weights went open that day, and the model joined Kimi Code.²²2Moonshot AI, Kimi Code changelog, entry of 12 June 2026. https://www.kimi.com/code/docs/en/kimi-code/whats-new.html Moonshot reports about 30% fewer thinking tokens than K2.6.¹¹

Kimi Code Bench v2 is Moonshot's in-house benchmark, with tasks in more than 10 programming languages. There the score went from 50.9 to 62.0. Both models ran with thinking on in Kimi Code CLI at a 262,144-token context.¹¹ GPT-5.5 scores 69.0 in the same table.¹¹

Program Bench asks an agent to rebuild a program from its binary and its documentation alone. Its 200 tasks run from small command-line tools to FFmpeg and SQLite. K2.7 Code reaches 53.6, against 48.3 for K2.6.¹¹ On MCP Atlas, with a budget of 100 tool calls, it moved from 69.4 to 76.0.¹¹

Thinking is always on, and so is preserved thinking. There is no instant mode.¹¹

Running it

The licence is K2.6's Modified MIT with a different name to display. A product above 100 million monthly active users, or US$20 million of monthly revenue, must show "Kimi K2.7 Code".³³3Moonshot AI, Modified MIT License of Kimi K2.7 Code. https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/LICENSE

Like K2.6, it ships in native INT4.¹¹ The card recommends vLLM and SGLang, and names Kimi Code CLI as the agent framework the model works best with.¹¹ It accepts video too, as an experiment, on Moonshot's own API only.¹¹

The API id is kimi-k2.7-code, with 262,144 tokens of context.⁴⁴4Kimi API Platform, chat model pricing, per million tokens, as listed on 29 September 2026. HighSpeed prices: $0.38 on a cache hit, $1.90 on a miss, $8.00 for output. Arithmetic: 0.38 / 0.19 = 1.90 / 0.95 = 8.00 / 4.00 = 2. https://platform.kimi.ai/docs/pricing/chat Cache misses cost $0.95 per million input tokens and hits $0.19. Output costs $4.00.⁴⁴ A HighSpeed version doubles every rate.⁴⁴