Back to AI Notes

Claude Cowork launches, Kimi K2.5 goes open-source

Anthropic shipped Cowork - Claude Code for everyone. And Moonshot AI open-sourced Kimi K2.5, a 1T parameter multimodal model that can chain hundreds of tool calls on its own.

Claude Cowork

link

Anthropic

Claude Code but for non-developers. Give Claude access to a folder on your computer, describe what you want done, and it makes a plan and works through it. Reads, edits, creates files, loops you in on progress, asks when it's stuck.

Think: "organize these 200 client files into folders by date and generate a summary spreadsheet." Tasks that would take hours but are straightforward to describe.

Connects to Google Drive, Gmail, DocuSign, and FactSet. Computer use performance jumped from under 15% to 72.5% on OSWorld - near-human on navigating spreadsheets, filling forms, managing files.

Everything runs locally on your machine - no data sent to external servers. Available on Pro ($20/month) and Max ($100/month).

Kimi K2.5

link

Moonshot AI

Open-source 1 trillion parameter MoE model from China, released January 27. 32 billion active parameters, 128K context, weights fully open under modified MIT license. K2.5 adds native multimodal capabilities - trained on 15 trillion mixed visual and text tokens.

The thing that sets it apart: it's built for agent workflows. Most open-source models are good at chat but fall apart when they need to actually do things. Kimi K2 can chain 200 to 300 sequential tool calls without human intervention. That's not a typo. Most models lose the plot after 10-15.

Kimi K2 Thinking is the reasoning variant - step-by-step reasoning with dynamic tool invocation. Like o1-style reasoning but with tool use baked in, and it's open-source.

The big draw is running a frontier-capable agent model on your own infra. No API dependency, no rate limits, no data leaving your network. For companies that can't send code to external APIs, this is huge.