Kimi rolls out K2.6 with Agent Swarm for large coding tasks

Kimi has unveiled version K2.6 of its artificial intelligence platform, introducing an Agent Swarm capability designed to break down large coding jobs into parallel tasks handled by cooperating AI workers. The Beijing-headquartered developer said the upgrade also turns uploaded documents into reusable skills, letting teams convert internal manuals and references into building blocks for automated workflows.

The release comes as Chinese AI specialists sharpen their focus on software engineering and agent-based automation, areas where domestic platforms are competing more aggressively with US-based large language model providers. Kimi, operated by Moonshot AI, has positioned K2.6 as a tool for full-stack website generation and document-driven workflows that can be reused across projects.

Alongside Agent Swarm, Kimi is previewing Claw Groups, a feature that allows multiple agents to collaborate on shared tasks. The company describes it as a way to simulate a development team inside a single prompt, with specialised agents handling coding, review, and testing in parallel before consolidating their output. Early promotional materials show the system writing front-end interfaces, configuring back-end services, and validating outputs without manual hand-offs between separate chat sessions.

For business users, the appeal lies in collapsing multi-step engineering projects into a single AI-driven cycle. Kimi markets the platform as a way to ship full-stack websites from natural-language prompts, with the model handling architecture decisions, database schema, and component styling. The document-to-skill conversion is pitched at enterprise teams that want to encode proprietary processes into the agent stack, reducing the time spent re-prompting the model on recurring jobs.

The K2.6 release lands against a backdrop of rapid iteration across China’s foundation model sector. Local providers have been stepping up the rollout of coding-oriented tools as overseas demand for AI-assisted development grows and as multinational cloud platforms localise their offerings for Chinese users. Kimi’s decision to bundle agent collaboration features rather than rely on single-threaded chat reflects a broader industry shift toward task-level decomposition, where work is distributed across specialised models that communicate through shared memory and tool calls.

Kimi said the new model is available through its web interface at kimi.com, where prospective users can trial Agent Swarm and request access to the Claw Groups preview. The platform competes in an increasingly crowded field that includes domestic rivals such as DeepSeek, Zhipu AI, and stepfun, all of whom have pushed coding-optimised models and multi-agent frameworks over the past year.

Found a mistake? See our corrections policy. Have a tip? Contact the editor.