Kimi K2.7 Code sharpens Moonshot's open-weight coding pitch

Moonshot AI has released Kimi K2.7 Code, a coding-focused successor to Kimi K2.6 aimed at long-context software engineering and tool-use workloads. The official Kimi API documentation now describes K2.7 Code as the company’s strongest coding model, while the Hugging Face model card says it improves long-horizon engineering performance and reduces thinking-token use by roughly 30% compared with K2.6.

The efficiency claim is central to the release. Agentic coding sessions spend tokens on planning, tool calls, retries, context management and intermediate reasoning before they produce a patch. A model that finishes more tasks with fewer reasoning tokens can change the cost profile even if headline input and output prices rise.

Kimi K2.7 Code remains a Mixture-of-Experts model with 1 trillion total parameters and 32 billion activated per token. The model card lists 61 layers, 384 experts, eight selected experts per token and one shared expert. Context length is 262,144 tokens, matching the 256K positioning from K2.6, and the release includes native image and video input through MoonViT.

The API version is deliberately constrained. K2.7 Code forces thinking mode, uses a fixed temperature of 1.0 and top-p of 0.95, and supports OpenAI-style chat completions through Moonshot’s API. The docs warn that multi-step tool calling must preserve the assistant message’s reasoning_content in context. That requirement may sound like a protocol detail, but it affects any team trying to drop the model into an existing agent loop.

Pricing is higher than K2.6’s headline rates while staying aggressive for a long-context coding model. Moonshot lists $0.19 per million input tokens for cache hits, $0.95 per million input tokens for cache misses and $4.00 per million output tokens. The useful comparison for engineering teams is completion cost across real tasks, including failed loops, rather than prompt cost in isolation.

The benchmark table shows clear improvement over K2.6, though the evidence should be read carefully. Moonshot does not publish SWE-Bench Verified or SWE-Bench Pro numbers for K2.7 Code in the same way that made K2.6 easy to compare with Claude and GPT models. Instead, the model card emphasises Kimi Code Bench v2, Program Bench, MLS Bench Lite, MCP Atlas, MCPMark Verified and Kimi Claw 24/7 Bench.

On Kimi Code Bench v2, K2.7 Code scores 62.0, up from 50.9 for K2.6. GPT-5.5 is listed at 69.0 and Claude Opus 4.8 at 67.4. On Program Bench, K2.7 reaches 53.6 versus 48.3 for K2.6, behind GPT-5.5 at 69.1 and Opus 4.8 at 63.8. The strongest relative movement appears in agentic and MCP-style tasks: K2.7 rises to 76.0 on MCP Atlas and 81.1 on MCPMark Verified, with the latter above Opus 4.8’s listed 76.4 but behind GPT-5.5’s 92.9.

Those figures support a narrower interpretation than a general frontier-model claim. K2.7 Code looks designed for repository exploration, patch generation, refactoring, test-writing and tool-heavy coding chores where long context and cost control count. The self-reported and model-provider benchmark elements mean teams should validate it against their own repositories before treating the scores as procurement evidence.

The open-weight release gives the model a different route into production from closed API systems. The Hugging Face card uses a Modified MIT licence with attribution requirements for commercial products above 100 million monthly active users or $20 million in monthly revenue. Deployment support through vLLM, SGLang and KTransformers, plus native INT4 quantisation, will interest teams that can justify the infrastructure for a 1T-parameter MoE model.

K2.7 Code is best understood as a specialised option in the coding stack. It gives Moonshot a stronger claim in agentic software work, particularly where cost and long-context execution dominate. The strongest closed models will still be preferred for ambiguous architecture, security-sensitive changes and final review, but K2.7 gives teams a credible open-weight model to test in the middle of the workflow.

Sources: Moonshot Kimi K2.7 Code docs, Kimi K2.7 Code model card, Kimi K2.7 Code pricing, Moonshot benchmark guidance.