What is Kimi K2.7 Code?
Kimi K2.7 Code is a coding-focused model from Moonshot AI, published on Hugging Face in June 2026 and built on Kimi K2.6. It is a mixture-of-experts model with 1 trillion total parameters and about 32 billion active per token, 384 experts with 8 chosen each step plus one shared, across 61 layers. The weights ship in native INT4 quantisation, the same method as Kimi K2 Thinking.
It reads text and images and writes text; video input is marked experimental and only works through Moonshot's own API. Context is 256K tokens. The model always runs in thinking mode and keeps its reasoning across turns, which cannot be turned off, but it is tuned to use about 30% fewer reasoning tokens than K2.6. The licence is a Modified MIT License covering both code and weights.
Is Kimi K2.7 Code worth it?
Moonshot's table shows a real generation-on-generation gain. K2.7 Code beats Kimi K2.6 on every benchmark listed: 62.0 against 50.9 on Moonshot's own Kimi Code Bench v2, 46.9 against 42.9 on its agentic Kimi Claw 24/7 Bench, and 81.1 against 72.8 on MCP Mark Verified. On that last test it even passes Claude Opus 4.8, which scores 76.4.
Everywhere else the closed models stay ahead. GPT-5.5 reaches 69.0 and Opus 4.8 67.4 on Kimi Code Bench v2 against K2.7 Code's 62.0, and both lead on MCP Atlas and the agentic bench. Two of the six headline tests, Kimi Code Bench and Kimi Claw 24/7 Bench, are Moonshot's own, so read the comparison as a vendor's framing rather than a neutral ranking. Notably, the card never reports SWE-bench, the test most buyers compare against.
The design is opinionated. Thinking and preserved reasoning are both always on, with no instant mode, so you cannot run it as a quick, cheap completer; the trade is that it uses roughly 30% less thinking than K2.6 for the same work. Because it shares the architecture of K2.5 and K2.6, existing vLLM, SGLang and KTransformers setups serve it unchanged, but a 1T-parameter model in INT4 still needs serious memory, which pushes most teams to the hosted API.
The catch: it is already a side branch. Moonshot's Kimi K3, published in July 2026, is the newer flagship, so K2.7 Code is best read as a coding-tuned fork of the K2 line rather than the model to standardise on. The permissive licence and the agentic tool-calling are the reasons to keep it in view.
Our take: if you are self-hosting the K2 family for coding agents and want the strongest open option in that line, this is it; for a fresh choice, weigh it against Kimi K3 and the hosted frontier before committing the hardware.
Where it earns its place
A platform team already serving Kimi K2.6 for coding agents
Upgrade to a stronger model without changing the serving stack
Swap in K2.7 Code, which shares the K2.5/K2.6 architecture and runs on the same vLLM, SGLang or KTransformers setup. Expect a step up across the coding benchmarks and roughly 30% less thinking per task, trimming token spend.
A company building an autonomous coding agent on open weights
Run long, multi-step software tasks with tool calls
Use the always-on thinking and interleaved multi-step tool calling to drive repository-level work through the Kimi Code CLI or an OpenAI-compatible server, accepting that you cannot switch to a cheaper instant mode for trivial calls.
A developer comparing open coding models under a permissive licence
Judge it against Qwen and DeepSeek before committing hardware
Weigh the Modified MIT terms and the agentic scores against a dense alternative like Qwen3.8-27B that fits one GPU; K2.7 Code's 1T INT4 weights need far more memory for its higher ceiling.
Before you commit
- Setup
- Weights on Hugging Face in native INT4; served with vLLM, SGLang or KTransformers and transformers >=4.57.1,<5.0.0. The same deployment as Kimi K2.5/K2.6 works. Thinking and preserve_thinking are always on.
- Integrations
- Kimi Code CLI and an OpenAI-compatible API; interleaved thinking with multi-step tool calls as in Kimi K2 Thinking; MoonViT vision encoder for image input.
- Data and privacy
- Modified MIT License on code and weights, so self-hosting keeps prompts on your own infrastructure. Hosted use through Moonshot's Kimi API follows Moonshot's terms, which the card does not restate.
- Cost at scale
- Self-hosted, the cost is the memory and compute for a 1T-parameter MoE with 32B active in INT4; the hosted Kimi API and Kimi Code plan are the practical route for teams without that hardware.
- Limits
- 256K context; no instant mode, so every call pays for thinking; video input is experimental and API-only; no SWE-bench figure published; superseded as a flagship by Kimi K3.
What’s good
- Beats Kimi K2.6 on every published benchmark
- Passes Claude Opus 4.8 on MCP Mark Verified (81.1 to 76.4)
- Permissive Modified MIT licence on code and weights
- About 30% fewer reasoning tokens than K2.6
What’s not
- Trails GPT-5.5 and Opus 4.8 on most tests
- Two of six headline benchmarks are Moonshot's own; no SWE-bench
- Always-on thinking, no quick instant mode
- Superseded as a flagship by Kimi K3
Best for
Teams self-hosting the Kimi K2 line who want the strongest open, agentic coding model in that family and can run a 1T-parameter INT4 model.
Kimi K2.7 Code alternatives
Kimi K2.7 Code: common questions
- Is Kimi K2.7 Code worth it?
- A capable open coding model that clearly beats its own predecessor and still trails the closed frontier, with an awkwardly specialised "always thinking" design. We rate it 4.1 out of 5.
- How much does Kimi K2.7 Code cost?
- Free weights under a Modified MIT License. Hosted on Moonshot's Kimi API and Kimi Code plan; per-token pricing is not stated on the model card.
- What are the downsides of Kimi K2.7 Code?
- Trails GPT-5.5 and Opus 4.8 on most tests; two of six headline benchmarks are Moonshot's own; no SWE-bench; always-on thinking, no quick instant mode; superseded as a flagship by Kimi K3.
- Who is Kimi K2.7 Code best for?
- Teams self-hosting the Kimi K2 line who want the strongest open, agentic coding model in that family and can run a 1T-parameter INT4 model.
- How much does Kimi K2.7 Code cost for a team or in production?
- Self-hosted, the cost is the memory and compute for a 1T-parameter MoE with 32B active in INT4; the hosted Kimi API and Kimi Code plan are the practical route for teams without that hardware.
- Is Kimi K2.7 Code safe for company data?
- Modified MIT License on code and weights, so self-hosting keeps prompts on your own infrastructure. Hosted use through Moonshot's Kimi API follows Moonshot's terms, which the card does not restate.
- What are the best Kimi K2.7 Code alternatives?
- The closest alternatives we have reviewed are Kimi K3, Qwen3.8-27B and DeepSeek-V4.1-Flash.
Answers checked Nov 14, 2026.
Make Kimi K2.7 Code? Show this rating on your site
