Kimi K2.7 Code

An August 7, 2026 engineering write-up from Alphamatch documents Cloudflare's production deployment of three optimizations, co-developed with the open-source SGLang inference framework, for Kimi K2.6, Kimi K2.7 Code, and GLM 5.2 on Workers AI. The techniques combine KV cache quantization from BF16 to FP8 (roughly doubl Measured on Kimi K2.6 decoding, FP8 sustains 2,192 tok/s at 64 concurrent requests where BF16 runs out of memory, yielding a reported 41% throughput improvement and 30% lower cost per token at peak FP8 concurrency with no measurable accuracy loss. For developers using @cf/moonshotai/kimi-k2.7-code on Cloudflare Workers

Discover how Moonshot AI’s Kimi K2.7 update is revolutionizing open coding models in China, driving innovation and transforming the AI landscape.

Kimi K2.7 Code launches with stronger coding performance, lower token usage, and improved agentic workflows. Here's what developers need to know.

Kimi K2.7-Code, Moonshot AI’s open-weight coding model, claims 30% fewer AI reasoning tokens than its predecessor and a new six-times-faster HighSpeed Mode — but no independent benchmark submission

Kimi K2.7 Code is Moonshot AI's open-weight coding-specialised model released on Hugging Face on June 12, 2026, under a Modified MIT license. According to the codersera guide drawing on the official model card, it uses a 1-trillion-parameter Mixture-of-Experts architecture with 32B active parameters per token, 384 expe The guide reports that Kimi K2.7 Code is tuned for long-horizon agentic software engineering and uses roughly 30% fewer reasoning tokens than Kimi K2.6, with the API model ID kimi-k2.7-code and pricing of $0.95 per million input tokens and $4.00 per million output tokens. The author notes that as of the June 12, 2026 p

Cloudflare's Workers AI changelog dated June 12, 2026 announces the general availability of @cf/moonshotai/kimi-k2.7-code on the platform, describing it as a frontier-scale 1T parameter MoE model optimized for coding with a 262.1k token context window, vision input, multi-turn tool calling, and reasoning support. The s For developers this means the Kimi K2.7 Code model can be invoked through Cloudflare's Workers AI runtime using the @cf/moonshotai/kimi-k2.7-code identifier, gaining access to one of the largest open-weight coding models without self-hosting. The changelog recommends reading the model-specific entry before migration, a

Artificial Analysis places Kimi K2.7 Code in the leading intelligence band of its open-weights class, scoring 26 on its Intelligence Index versus a class median of 18, while generating 120M tokens (fairly concise compared with the 140M median). The platform rates the model 4 out of 4 for intelligence and 3 out of 4 for The Artificial Analysis page confirms Kimi K2.7 Code is a reasoning-capable open-weights model with 1T total parameters, 32B active per token, a 256K-token context window, and support for text, image, and video input with text output, priced at $0.95 per 1M input tokens and $4.00 per 1M output tokens with an 80% cache

Moonshot AI released Kimi K2.7 Code on June 12, 2026 as an open-weight, coding-focused agentic model built directly on Kimi K2.6. It uses a 1 trillion-parameter Mixture-of-Experts architecture with 32 billion active parameters and a 256K (262,144) context window, supports text, image, and video inputs, and runs always- The release uses a two-track strategy: full open weights on Hugging Face under a Modified MIT license with day-one vLLM and SGLang support, paired with Kimi Code, a first-party terminal-native coding agent with its own subscription tiers. The article also covers pricing and access channels, making it a practical entry