Published:
Moonshot AI upgraded Kimi Work — its local desktop AI agent for knowledge workers — to run on the Kimi K3 model on July 17, 2026, when Moonshot launched K3 across its full product suite, making Kimi Work the first publicly available local AI agent backed by a 2.8-trillion-parameter frontier model. Kimi Work runs on macOS (Apple Silicon) and Windows, orchestrates up to 300 parallel sub-agents on the user’s device, and controls the user’s real browser through a WebBridge extension — without uploading local files to a vendor cloud.
What Is Kimi Work and What Does It Do?
Kimi Work is a local desktop agent that reads files stored on the user’s machine, executes tasks in a real logged-in browser session, and runs scheduled background jobs using a built-in cron scheduler. Moonshot AI, based in Beijing, first announced Kimi Work on June 12, 2026, according to MarkTechPost. The product is available in public beta at kimi.com/products/kimi-work. Kimi Work can generate finished output in PowerPoint or Excel format from completed research tasks, and pre-integrates market data from A-share, Hong Kong, and US equity markets.
How Does the 300-Agent Swarm Work?
Kimi Work deploys up to 300 sub-agents running in parallel on the user’s local machine, coordinated by a top-level agent that decomposes an instruction into smaller tasks and assigns each to a sub-agent. The sub-agent count — documented at up to 4,000 coordinated steps in the original June announcement — was first reported for the K2.6 version; the K3 sub-agent capacity is not separately confirmed. Businesses evaluating the best AI tools for business productivity will find Kimi Work relevant for research, document generation, and browser automation tasks that require accessing private local data without cloud uploads.
Kimi Work’s WebBridge browser extension drives the user’s own browser — including logged-in sessions and first-party cookies — rather than a sandboxed browser instance. This lets the agent search, scroll, extract structured data, and submit forms as the user. An “ask before acting” approval gate covers file write operations; a “YOLO mode” can disable this gate for automated pipelines.
How Does a Local AI Agent Differ from Cloud AI Tools?
Kimi Work’s local architecture separates it from cloud-based agent platforms such as ChatGPT Work, which runs tasks in OpenAI’s servers and requires users to upload files or grant OAuth access to third-party services. Kimi Work’s agent reads local files directly and accesses the user’s real browser session — no file upload and no third-party OAuth required. The trade-off: availability is unconfirmed for users outside Asia, and the beta’s subscription pricing has not been disclosed.
One technical clarification on the “local” framing: Kimi Work orchestrates agents locally, but the K3 model itself (2.8 trillion parameters) almost certainly runs on Moonshot’s servers, not on the user’s device. “Local” refers to the agent’s access layer — files and browser — not to where the model weights execute.
What Did the K3 Upgrade Add to Kimi Work?
K3 replaces K2.6 as Kimi Work’s underlying model. K2.6 is a 32-billion-active-parameter mixture-of-experts model with a 256K token context window; K3 runs 2.8 trillion parameters and extends context to 1 million tokens. The larger context window allows Kimi Work to hold longer documents, more browser-extracted content, and larger code files in a single working session without truncation. Businesses comparing the best AI agents for business tasks can treat K3’s 1M-token context as the practical differentiator — most competing local agents top out at 128K–256K.
Kimi Work is the most capable local AI agent most Western business users have not yet evaluated. A 300-agent swarm reading your files, running your browser, and executing overnight scheduled jobs — powered by the world’s largest open-weight model — is a materially different product category from cloud-hosted AI assistants. The privacy-control angle (no file upload, no vendor cloud, real browser sessions) addresses the exact objection that has kept enterprises from deploying cloud AI agents on sensitive workflows. The caveat: regional availability outside Asia is unconfirmed, and “local orchestration” is not the same as “model runs on your hardware.” Watch for the open-weight K3 release (promised by July 27, 2026) to clarify whether on-device inference becomes viable.
Related Coverage
- Meta entered the same enterprise agent API market with Meta’s agentic AI models via Muse Spark 1.1 — a cloud-based alternative to the local-first approach Kimi Work takes.
For Context: Kimi K3 and Moonshot AI on WithO2
- Kimi K3 Is Here: Open-Weight Model That Rivals Claude Fable 5 — our coverage of the K3 model launch on July 17, 2026, including benchmark results and the open-weight release timeline.
