Moonshot AI's Kimi Work Unleashes 300 Parallel Agents on Your Desktop

Kimi Work is Moonshot AI's new desktop application that turns your local machine into a multi-agent command center. Rather than a chat window that hands you text to act on, Kimi Work is designed to execute tasks end-to-end: browsing the web, reading and writing local files, running scheduled jobs overnight, and delivering finished documents straight to your desktop.

What's actually shipping

The announcement bundles four distinct capabilities into one desktop app:

• Agent Swarm: Up to 300 specialized sub-agents running in parallel on your local machine, each with its own tools and context. Think of it as spawning a small team of AI workers that divide a complex task and execute their pieces simultaneously, rather than one agent plodding through steps sequentially.
• WebBridge: A Chrome/Edge browser extension that lets the agent click, scroll, type, and extract data from real websites inside your existing browser session. Everything runs locally, so your login sessions and page content never leave your device.
• Finance data tool calls: Native integrations with Yahoo Finance, World Bank, and Binance , no API keys to configure. You can ask for earnings reports, macro data, or crypto prices in plain language and get structured results back.
• Memory system: A persistent diary of your preferences, past decisions, and context that accumulates across sessions so the agent doesn't start from scratch every time.

On top of those, a built-in Cron scheduler lets you set tasks to run on a timer , daily briefings at 6 AM, overnight data processing, recurring reports , with an option to keep the machine awake so nothing is missed.

The engine underneath

Kimi Work runs on Kimi K2.6, Moonshot AI's latest open-weight model. It's a 1-trillion-parameter Mixture-of-Experts model with only 32B parameters activated per token, released fully open-source under a Modified MIT License. The MoE design (Mixture-of-Experts is an architecture where only a fraction of the model's parameters are activated for any given input, keeping inference cost low while maintaining a large knowledge capacity) is what makes running 300 agents locally even plausible , each agent call is cheap because the model activates only ~32B parameters at a time.

The Agent Swarm architecture parallelizes complex tasks across many specialized sub-agents rather than relying on a single, deeper reasoning chain. It scales horizontally to 300 sub-agents executing across 4,000 coordinated steps simultaneously , a substantial expansion from K2.5's 100 sub-agents and 1,500 steps.

When Moonshot AI introduced Agent Swarm in January 2026 alongside Kimi K2.5, it was a genuine architectural bet: absorb multi-agent orchestration directly into the model, let it self-decompose tasks, spawn sub-agents on the fly, and synthesize results , no LangGraph, no CrewAI, no manually-configured workflow templates required.

Benchmarks worth noting

The underlying K2.6 model posts strong numbers on agentic benchmarks, though it's worth noting these are largely vendor-reported figures:

• K2.6 leads all frontier models on HLE-Full with tools (54.0), outperforming GPT-5.4 (52.1), Claude Opus 4.6 (53.0), and Gemini 3.1 Pro (51.4).
• BrowseComp in Agent Swarm mode scores 86.3, up from 78.4 on K2.5. DeepSearchQA F1 reaches 92.5.
• On SWE-Bench Verified it scores 80.2%, and 58.6% on SWE-Bench Pro.

What it's actually good for

The finance angle is the most concrete use case at launch. Kimi Work comes pre-wired to pull live market data from Yahoo Finance, World Bank macroeconomic datasets, and Binance. A single prompt like "give me a competitive analysis of TSLA vs RIVN with macro context" can trigger a swarm that pulls earnings data, economic indicators, and price history in parallel, then assembles a finished Excel or PowerPoint file.

Beyond finance, the swarm architecture shines on tasks that decompose naturally into parallel workstreams:

• Deep research reports (literature reviews, market surveys, competitive landscapes)
• Batch document processing , summarizing a folder of PDFs, extracting structured data from contracts
• Overnight scheduled jobs , daily briefings, automated data pulls, recurring report generation
• Web automation , price monitoring, form filling, cross-site data extraction via WebBridge

The most eye-catching demo is a 104-page literature review produced in one shot , a 10,000-word output downloadable as Word, PDF, PPT, or Excel. That's the kind of task that previously required either hours of manual work or a carefully orchestrated custom pipeline.

Where it falls short

The swarm architecture has a real failure mode: decomposition quality is the first place a swarm can fail. If the coordinator's understanding of what the task requires is wrong , because the initial prompt is ambiguous, or because the task is genuinely novel , the subtask assignments will be wrong from the start, and sub-agents will work efficiently in the wrong direction.

As of now, all published benchmark results for K2.6's Agent Swarm , including the BrowseComp 86.3% and DeepSearchQA 92.5% figures , come from Moonshot AI's own model card. Independent third-party replication is in progress. VentureBeat's reporting separately quotes a practitioner noting that "orchestration is still fragile... it feels more like a product and training problem than something you can solve by writing a sufficiently stern prompt."

The architecture is not the right choice for teams that need to explain their agent decisions, tune their orchestration logic, or maintain strict cost control across variable-complexity swarm runs. If you need full auditability of every agent step, a hand-rolled LangGraph pipeline will give you more control.

Getting started and what it costs

Kimi Work is available now as a free download for macOS (Apple Silicon) and Windows. The desktop app itself has no separate price tag , it runs on your Kimi account. Membership starts at $19/month (Moderato) and gives you K2.6 inside the Kimi chat interface with agent credits. Higher tiers , Allegretto ($39), Allegro ($99), and Vivace ($199) , unlock Agent Swarm with up to 300 parallel sub-agents, more Kimi Code credits, and larger data quotas.

For developers who want to wire K2.6 into their own pipelines directly, the API is fully OpenAI-compatible , swap in model: "kimi-k2.6" and you're running the latest model in any existing workflow. API pricing sits at $0.95 per million input tokens and $4.00 per million output tokens for K2.6, with automatic caching dropping cached input to $0.10–$0.16/M tokens.

Setting up WebBridge takes about a minute: download the desktop app, install the WebBridge browser extension, switch to Kimi Work mode, and send a prompt. The extension uses Chrome DevTools Protocol to drive your existing browser, which means it operates within your logged-in sessions without needing to re-authenticate to every site.

The bigger picture

Kimi Work is Moonshot AI's answer to a question the whole industry is circling: what does an AI agent look like when it's not a chatbot with tool access bolted on, but a system designed from the ground up to execute work autonomously? K2.6 continues the trajectory set by Kimi K2 in mid-2025: an openly licensed Chinese model that trades blows with the leading US closed systems on coding and agentic benchmarks, released with weights and deployment recipes rather than an API-only posture.

Cursor recently confirmed that its Composer 2 coding model was built on top of Kimi K2.5, which gives some signal about the underlying model quality beyond vendor benchmarks. The desktop app wraps that same model family in a consumer-facing product aimed squarely at knowledge workers who want results delivered to their desktop, not a prompt interface they have to babysit.

The 300-agent ceiling, the finance data integrations, and the Cron scheduler together represent a bet that the next wave of AI adoption isn't about smarter chat , it's about autonomous background workers that run while you sleep and hand you finished work in the morning.