Moonshot Open-Sources Kimi-K2.7-Code, Says It Uses 30% Fewer Reasoning Tokens
Moonshot released and open-sourced Kimi-K2.7-Code, its latest coding model, saying it improves on K2.6 by 21.8% on Kimi Code Bench v2, 11.0% on Program Bench and 31.5% on MLS Bench Lite while using 30% fewer reasoning tokens. The company said the model also improves instruction following and end-to-end success rates on longer coding tasks, and made it available through Kimi API and Kimi Code with weights and code published separately.
The vLLM project said Kimi-K2.7-Code is already supported in vLLM using the same deployment as K2.6, and described it as a 1-trillion-parameter mixture-of-experts model with 32 billion active parameters per token and a 256,000-token context window. Moonshot said a 6x High-Speed Mode is coming soon.
From the sources (21 posts)
@kimi_moonshot🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less
@kimi_moonshot🎸 We're also launching the Kimi Code Beta Program today. Apply now if you'd like to try upcoming models and features before public release 👉
@stevibeKimi K2.7-Code!
@kimi_moonshot🔗 Weights & code:
@teortaxestexI want to see this compared with Composer 2.5 Like, really hard Cursor has a ton of proprietary data, a large head start, and threw a Colossus at RLing Kimi K2.5 checkpoint. What is the gap now?
@vanstriendanielRT @Kimi_Moonshot: 🔗 Weights & code:
@crystalsssupless overthinking 👀
@teortaxestexHUGE @htihle and others will get to test this soon I hope
@zephyr_z9Great work from KIMI
@eliebakouchRT @Kimi_Moonshot: 🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over…
@code_starIt would be really cool if the top Chinese labs could properly benchmark and evaluate Fable/Mythos to show how it compares to their releases. I guess that isn’t possible though.
@thezachmuellerRT @Kimi_Moonshot: 🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over…
@kimmonismusMoonshot just released Kimi-K2.7 code, a huge upgrade to Kimi-K2.6! Big jump over K2.6: +21.8% on Kimi Code Bench v2 +11.0% on Program Bench +31.5% on MLS Bench Lite It also uses 30% fewer reasoning tokens, follows instructions better, and
@andrewcurran_RT @Kimi_Moonshot: 🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over…
@nrehiew_I think K2.6 is the third best model on the planet - albeit one of the slowest. With the overthinking fixed and overall performance improvements, can’t wait to use this
@vllm_project🎉 Congrats to @Kimi_Moonshot on Kimi K2.7-Code, a coding-focused agentic model built on K2.6. ✨ 1T-parameter Mixture-of-Experts, 32B active per token ✨ MLA attention with a 256K-token context window ✨ ~30% fewer thinking tokens than K2.6 f
@clementdelangueRT @Kimi_Moonshot: 🔗 Weights & code:
@kimi_moonshotRT @vllm_project: 🎉 Congrats to @Kimi_Moonshot on Kimi K2.7-Code, a coding-focused agentic model built on K2.6. ✨ 1T-parameter Mixture-of-…
@matthewbermanNearly frontier open source coding model. Congrats to the Kimi team. I would love to see this tested against DeepSWE
@thezachmuellerRT @elliotarledge: I benchmarked Kimi K2.7-Code (1T MoE, coding-specialized, just dropped) on KernelBench-Hard, against its general predece…
@threepointoneRT @michellechen: k2.7 code live on workers ai