Kwaipilot Releases KAT-Coder V2.5 Dev, an Agentic MoE Coder

Kwaipilot, the developer-tools group inside Kuaishou, has published KAT-Coder V2.5 Dev on Hugging Face. The model is aimed squarely at agentic coding — the kind of multi-step, tool-using workflows where an assistant plans, edits, and iterates on code rather than simply completing a single prompt.

Under the hood, KAT-Coder V2.5 Dev is a mixture-of-experts (MoE) model built on the Qwen3.5 MoE architecture. MoE designs route each token through a subset of specialized expert networks, which lets a model carry a large total parameter count while keeping the compute cost of any single forward pass lower than a comparably sized dense model. That trade-off is increasingly popular for code models, where breadth of knowledge matters but latency and cost still bite.

Why it matters

The coding-model space has become one of the most competitive corners of open-weights AI, and Kuaishou's entry adds another well-resourced backer to the field.

• It leans on the widely adopted Qwen MoE lineage, which lowers the barrier for tooling and integration.
• Its focus on agentic workflows aligns with where developer tools are heading, from single-shot completion toward autonomous task execution.
• As an openly published model, it gives teams an alternative to closed coding assistants for self-hosted or customized deployments.

Some key details remain unstated in the release, including exact parameter counts, active-expert configuration, and context length. The model is distributed under a custom license, so teams evaluating it for production should review the terms on the model page before building on top of it.