Kimi K2.7 Code Highspeed

The kimi-k2.7-code-highspeed variant is tuned for roughly 180 tokens/sec (up to 260 tok/s in short contexts), delivering roughly six times the throughput of the standard endpoint. It ships alongside kimi-k2.7-code, both released under a Modified MIT license that covers the model weights themselves, making this a genuin A notable design constraint is that thinking mode cannot be disabled on these models — every request runs the full chain-of-thought regardless of caller preference, and the API errors if users try to override temperature, top_p, or the penalty parameters away from their fixed defaults. Moonshot positions this as a deli

A third-party AI roundup dated June 16, 2026 explicitly names "Kimi K2.7 Code HighSpeed" and frames it not as a separate model but as a speed-optimized mode of the open-source Kimi K2.7 Code. It reports the same headline figures of approximately 180 tok/s on median-length coding tasks and peaks of up to 260 tok/s in sh The article attributes the speed gains to speculative decoding combined with optimized kernel scheduling tailored specifically for code-completion patterns, where token distributions are more constrained than in general chat. It reports early-adopter feedback of noticeable latency reductions in interactive coding sessi

Moonshot AI's platform documentation introduces Kimi K2.7 Code HighSpeed as a high-speed variant of the Kimi K2.7 Code coding model. The HighSpeed edition shares the same underlying model but outputs at approximately 180 tokens/s, reaching up to 260 tokens/s in short-context scenarios, offering a faster coding experience than the standard variant. The docs describe Kimi K2.7 Code HighSpeed as supporting a 256K context window, improved instruction compliance, and long-horizon coding versus the K2.6 predecessor, with Moonshot reporting roughly 30% less overthinking and 10% better agentic performance in external benchmarks. The API is OpenAI-compatible, no non-thinking mode is offered, and the HighSpeed tier is noted to be resource-constrained while Moonshot gradually expands capacity.

An encyclopedia entry explicitly titled "Kimi K2.7 Code Model High-Speed Edition" attributes the model to Moonshot AI ("Dark Side of the Moon") and dates the HighSpeed launch to June 15, 2026, three days after the standard K2.7 Code release on June 12, 2026. It reports the same ~180 tokens/s typical output speed and up The entry describes the initial access rollout as limited to Kimi Code Beta Program members, API developers, and Kimi Business users, with a planned gradual expansion to Kimi members at the Allegretto level and above. It also notes that thinking mode must be enabled to achieve optimal performance with the model series.

Overchat's model directory page restates that Kimi K2.7 Code is a coding-focused open-weight flagship in Moonshot AI's Kimi K2 family, released on June 12, 2026 as the successor to K2.6 and shipped under a Modified MIT license with weights on Hugging Face at moonshotai/Kimi-K2.7-Code. It repeats the 1T-parameter MoE ar The same listing describes a separate kimi-k2.7-code-highspeed endpoint aimed at latency-sensitive agentic loops, citing the same ~180 tok/s typical and ~260 tok/s short-context throughput figures and a roughly sixfold speedup over the standard endpoint. It positions K2.7 Code as 5–12× cheaper per token than Claude Opu

AgentsFlare reports that in June 2026 it added four new models: GLM-5.2, Kimi K2.7-Code standard and high-speed variants, and the GA version of Gemini 3 Pro Image. The update wave is concentrated around two capability tracks, coding agents and image generation. On the coding track, Moonshot released Kimi K2.7-Code and The AgentsFlare update states that Kimi K2.7-Code improves coding capability and reduces reasoning-token usage without raising its base price, positioning it competitively against GLM-5.2's approach of approaching Claude Opus 4.8-level engineering performance at lower cost. The post also flags retirement and deprecatio