LM Studio’s Bionic Wants to Turn Your Local AI Models into an Actual Assistant

For a couple of years now, the pitch for running AI models on your own machine has been mostly about principle. It's private, it's free, it's yours. All true, and all a little abstract when the thing you actually got was a chat window that answered questions. LM Studio, the app a lot of people use to run open models locally, has been that chat window. Now it wants to be something you can put to work.

The company just launched Bionic, an app for Mac and Windows that it calls "the AI agent for getting real work done with open models." The key word there is agent. Instead of just talking back, Bionic is built to go do things: dig through your code, read your documents, and take actual steps toward finishing a task. And it does it on models you control, whether those run on your own hardware or on LM Studio's cloud.

Here's the setup. Bionic runs open models two ways. Small and medium ones run locally through LM Studio's runtime, the same way the app has always worked. When you need more horsepower, you can push the job to LM Studio Secure Cloud and tap bigger open-source models, all under a Zero Data Retention policy. That last part matters, because the whole reason people reach for open models is that they don't want their work sitting on someone else's server. Using the cloud does require an account with billing set up, so the local-first crowd should know the free part stops at your own silicon.

What can it actually do? Three things, mostly. There's a voice keyboard that works across your apps and transcribes offline using Mistral's Voxtral model, in multiple languages, so you can talk instead of type without shipping your audio anywhere. There's code work, where you point Bionic at a local folder and let models like GLM 5.2 and Kimi K2.7 Code investigate, edit, and debug, complete with inline diffs and agentic search through your project. And there's document work: it chews through PDFs, presentations, and spreadsheets inside a sandbox, with web search, summarization, and automatic checkpoints so you can roll back when it does something you didn't want.

Those checkpoints are the tell that LM Studio has been paying attention. Anyone who has handed a coding agent the keys knows the small panic of watching it confidently rewrite the wrong file. Building in a safe undo, and running document work in a sandbox, is the kind of unglamorous decision that separates a demo from a tool you'd trust with real files.

Who is this for? Not the person happy chatting with a local model for fun. It's for the developer who wants an agent that isn't beaming their codebase to a frontier lab, and the privacy-minded worker who wants document help without the terms-of-service roulette. The open-model world has spent a while proving it can match the big names on raw answers. The harder question was always whether it could do the work. Bionic is LM Studio betting the answer is finally yes.