Kimi K2.5

Lorphic's July 2026 explainer walks through the Kimi K2 model family, explicitly distinguishing K2, K2.5, K2.6, and K2.7 as architecturally related but with distinct capabilities and licensing. The piece confirms every K2-line model shares a 1-trillion-parameter MoE design with 32B active parameters, 384 routed experts The article situates Kimi K2.5 within the broader release timeline as an intermediate family member, useful for practitioners deciding which K2 variant to target. It frames the MoE architecture as the reason Kimi models can be priced competitively against dense models of comparable scale, since inference cost tracks ro

CoreWeave Inference achieves the highest output speed for newly-launched Kimi K2.7 Code and ranks in the most attractive price-performance quadrant., Stream

Kimi K2.7-Code claims 30% fewer thinking tokens and a drop-in API swap path, but independent benchmarks show kernel regressions and no DeepSWE submission.

CoreWeave Inference achieves the highest output speed for Kimi K2.6 and ranks in the most attractive price-performance quadrant.

Cursor's Composer 2.5 undercuts Opus 4.7 and GPT-5.5 on price, posts gains on Terminal-Bench and SWE-Bench, but real-world coding tests loom.

Open-source push continues with Kimi K2.6, but pressure to monetise is nudging some Chinese AI firms behind closed doors.

Kimi K2.5 brings advanced agentic reasoning, design-to-code capabilities, and production-ready performance to fully managed inference endpoints on Nscale. Follow our blog for more insights from Nscale.

The US AI startup Cursor has acknowledged that its newly introduced coding model Composer 2 is built on the Chinese open-source language model Kimi K2.5

Alibaba Cloud has launched a new Coding Plan featuring four open-source model API services: Qwen3.5, GLM-5, MiniMax M2.5, and Kimi K2.5. According to PANews, this initiative makes Alibaba Cloud the on

Moonshot AI released Kimi K2.5 in January 2026, an open-source model built on a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion activated per request. According to the Codecademy guide, it was trained on 15 trillion tokens that mixed visual and textual data from the start, letting vision The Codecademy article highlights Kimi K2.5's headline benchmark: 50.2% on Humanity's Last Exam at roughly 76% lower cost than Claude Opus 4.5, alongside claims of a 4.5x execution-time speedup from its parallel-agent approach. The Agent mode supports autonomous workflows with 200–300 tool calls. The piece positions K2