Moonshot AI has released Kimi K2.6, an open-weight AI model built to compete with leading frontier systems on coding and agentic benchmarks. The release puts a heavy emphasis on parallel work: K2.6 can coordinate up to 300 agents at the same time, turning one request into many specialized subtasks.
The company presents the model as comparable with GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro across several benchmark results. It is also being positioned not only as a chat model, but as a system for long-running tool use, coding, web work, and multi-agent production tasks.
What Moonshot AI is claiming for Kimi K2.6
According to Moonshot AI, K2.6 reaches top scores across several evaluations. The reported results include 54.0 on HLE with Tools, 58.6 on SWE-Bench Pro, and 83.2 on BrowseComp.
Those figures matter because they point to the kinds of work Moonshot AI wants K2.6 to be judged on: tool use, software engineering, and browsing-based tasks. Instead of describing the model mainly as a general chatbot, the release frames it around practical execution.
K2.6 can chain together more than 4,000 tool calls and run continuously for over twelve hours. Moonshot AI says that sustained operation applies in languages like Rust, Go, and Python, which makes coding a central part of the model’s pitch.
The broader claim is straightforward: Kimi K2.6 is designed to stand near GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on the kinds of benchmarks that matter for building, browsing, and acting with tools.
Agent Swarm is the headline feature
The release’s most distinctive feature is Agent Swarm. It can run up to 300 sub-agents at once, with each agent taking 4,000 steps.
In plain terms, Agent Swarm is meant to divide a large assignment into smaller jobs. The system then sends those jobs to specialized agents, rather than asking one model thread to handle everything sequentially.
Moonshot AI says the agents can combine skills such as web research, document analysis, and writing. The intended output is not just a draft or a partial answer, but finished materials from a single run.
Examples of those finished outputs include:
- documents
- websites
- slide decks
- spreadsheets
That makes the agent swarm approach important for a specific reason. It shifts the focus from answering questions to completing workflows. A request can become a coordinated process in which different agents gather information, analyze it, structure it, and produce a final asset.
Humans can join the workflow through claw groups
K2.6 also includes a preview feature called "claw groups." This feature allows multiple agents and humans to work together as a team.
In that setup, K2.6 handles coordination. It assigns work based on each agent’s strengths, and it can step in when an agent fails or gets stuck.
The source does not describe claw groups as a finished product. It calls the feature a preview, which means it should be understood as part of the direction Moonshot AI is taking with K2.6 rather than only a benchmark result.
Still, the concept is significant. Many AI workflows require both automation and human judgment. A system that can coordinate agents while allowing people to participate could make multi-step work easier to manage, especially when the task involves research, documents, and writing.
Website generation goes beyond the front end
Moonshot AI also says K2.6 can create complete websites from text prompts. The source specifically mentions animations and database connections, as well as the use of image and video generation tools to keep visuals consistent.
This is not framed as simple front-end generation alone. Moonshot AI says the model can handle basic full-stack tasks too.
The listed examples include user sign-ups, database operations, and session management. Those are practical parts of a working web application, not just visual layout or page styling.
For developers and teams evaluating Kimi K2.6, that claim places the model in a broader category than code completion. The promise is that the model can help assemble a working digital product from a prompt, including parts that connect the interface to application behavior.
License terms and availability
K2.6 is available under a modified MIT license. The source describes the license as allowing largely free use, with one specific condition for very large commercial deployments.
Anyone deploying the model in commercial products with more than 100 million monthly active users or over $20 million in monthly revenue has to visibly credit "Kimi K2.6" in the user interface.
The model is available in several forms. Users can access it on kimi.com in chat and agent mode, use it as a coding tool through Kimi Code, reach it via API, or download it as open source on Hugging Face.
That mix of access points is part of the release’s positioning. Kimi K2.6 is presented both as a product users can try directly and as an open-weight model that developers can download and deploy under the stated license terms.
The key takeaway is that Moonshot AI is pushing K2.6 around three connected ideas: open-weight access, strong coding and tool-use performance, and large-scale agent coordination. If those pieces work as described, Kimi K2.6 is less about a single chat session and more about turning AI into a managed production system for complex tasks.