This week, AI went to the "office". Linus Torvalds says he actually likes using AI, and his new kernel branch credits Claude Code for the code, the tests and even a draft of the commit message. Office suites are being rebuilt so agents can edit your documents. And Big Tech wants its agents inside every work app you use, including the office app.
The question I keep asking is whose desk the agent sits on. If it lives in someone else's cloud, so does your work. That's why the local and self-hosted options matter.
🧪 On my desk: openKylin 3
I've been testing openKylin 3.0. It ships with Kylin bot, an AI assistant that can change desktop settings for you. With openKylin's own hosted model, I could successfully make the agent change the system settings.
But my local and frontier models did not fare as well. When assigned a task, my Qwen model with Ollama replied that it changed system settings when it was not changed in reality.
I will be sharing all this in a video later this week. Stay tuned for that.
🔍 Discover AI tools
BetterOffice is an Apache-2.0 office suite you embed in your own web app, with DOCX, XLSX and PPTX editors built on Rust and WebAssembly. The interesting part is that an AI agent edits as a peer, and its changes show up as tracked changes you accept or reject.
ArtCraft, the folks behind those Adobe clones, has moved on to Microsoft Office. WordCraft, GridCraft and DeckCraft are Rust clones of Word, Excel and PowerPoint, built with Claude. The READMEs claim high parity with Office, but the numbers are self-reported and the apps are pre-alpha. I'll believe it when it is tested in real world, with real use cases.
Razorpay open sourced slashdocs, a self-hosted place where agents publish HTML reports and your team comments on them. The agent reads the comments and pushes the next version. It's MIT licensed and ships as a single Go binary.
Raven calls itself a "harness of harnesses." It splits a task into pieces and hands them to specialist agents, including Claude Code, Codex, OpenCode and Hermes. It's Apache-2.0 and pre-alpha (again). Sorry but so many new tools in alpha stage these days. But that's the joy of discovering new tools, right?
📡 Open Model News
Two big announcements this week, not that you can download them (yet).
Mistral Large 4, nicknamed "Le Chonk," is a 1-trillion-parameter multimodal MoE with roughly 50B parameters active. It's API-only in preview, and Mistral promises the weights by the end of October. No license has been announced yet.
Reflection AI's Beam is a 501B model with 23B active, aimed at coding and agentic work. The weights are due later this month under Apache 2.0. Reflection's benchmark claims haven't been independently verified yet.
Neither will fit on our regular machine anyway. If you want something small for coding agents, JetBrains released Mellum2.1 Thinking, a 12B model with about 2.5B parameters active.
👀 Big Tech Watch
At its Windows event, Microsoft pitched agents that run on your device when possible and in the cloud when not. Its MAI-Code-1.1-Flash runs locally as a 3-bit build with 256K context. Local calls are free, but the recommended RAM is 120GB or more (LOL).
Meanwhile, Google's new Gemini agent works across Workspace, Microsoft 365 and Slack for hours or days at a time. OpenAI launched always-on Dots agents in ChatGPT. Meta plugged its Muse agent into Slack, Shopify and QuickBooks. Everyone wants a seat at your desk.
🗂 AI Jargon: Tool calling
Remember Kylin bot claiming it changed the system setting? That was a tool calling failure.
A model can't touch your computer on its own. It can only produce text. Tool calling is how it asks the app to act: the app tells the model which tools exist, like set_volume, and the model replies with a structured request to use one. The app runs the tool and sends the result back.
Smaller models are often bad at this. They describe the action, or skip the call entirely and just say it's done. That's why tool support matters when you pick a model for an agent.
😂 Meme
An old meme that never really gets old.
⚡ Quick Tip: Check if your Ollama model can call tools
A quick check before you plug a local model into an agent like Hermes.
Run ollama show <model-name> and look at the Capabilities section. If tools isn't listed, the model wasn't trained for tool calling, and your agent will struggle no matter how you configure it.
Even if tools shows up, test it with a harmless action and verify the result yourself.
💡 If you have not subscribed to Local AI Weekly yet, you can subscribe from this page.
See you next week.
If this helped you, consider supporting It's FOSS
It's FOSS has been helping people use Linux for the past 14 years. Help us stay independent from big tech. Become a Plus member, enjoy ad-free reading and get 5 eBooks.
