Kimi K2.6 Launches on Huawei Cloud – Experience the New AI Model Today

On April 20, the Kimi K2.6 large language model was officially released and open‑sourced, bringing industry‑leading capabilities in code generation, long‑range task execution, and a scalable Agent cluster. Huawei Cloud completed targeted adaptations and optimizations, and its MaaS (Model‑as‑a‑Service) platform now offers a token‑based, one‑click Kimi K2.6 API for developers.

The model achieved top results on several benchmark suites, including the full‑scale “Humanity’s Last Exam” (Humanity's Last Exam), the software‑engineering focused SWE‑Bench Pro, and the deep‑search evaluation DeepSearchQA, where it outperformed competing systems.

Kimi K2.6 dramatically enhances autonomous Agent execution: it now supports up to 300 sub‑Agents running in parallel to complete 4,000 collaborative steps. Its long‑duration coding ability allows uninterrupted coding for 13 hours, producing or modifying more than 4,000 lines of code to develop and optimise complex systems.

Huawei Cloud leverages a KV‑Cache‑Aware intelligent scheduling technique combined with deep operator‑graph optimizations in the model architecture. This three‑layer system‑level optimization (scheduler → engine → operator) accelerates inference, especially for long‑sequence low‑latency scenarios, reducing time‑to‑first‑token (TTFT) by roughly 10 % compared with industry averages.

Through the MaaS platform, developers can obtain free, no‑deployment access to Kimi K2.6 via token services. The model is also integrated into Huawei Cloud CodeArts (code‑intelligent Agent) and OfficeClaw (office‑intelligent Agent), and can be accessed via the AgentArts development platform, OpenClaw, or Hermes Agent deployed on Huawei Cloud Flexus for rapid custom Agent construction.

Complete benchmark scores are published in the technical blog at https://www.kimi.com/blog/kimi-k2-6.