What it is, in one line: LM Studio Bionic is a standalone desktop agent — launched July 16, 2026 — that runs open models on your own machine to do real work: repo-aware coding with inline diffs, and document, PDF, deck, and spreadsheet work. When a job is too big for local hardware, it offloads to a Secure Cloud of frontier open weights with Zero Data Retention by default. If shipping code or client files to a hosted frontier API is your blocker, this is the local-first path.
It's a separate app from the original LM Studio chat client — the chat app was for talking to a model; Bionic is for handing a model a task and letting it act.
Who it's for#
Founders and small teams whose product or clients make "send this to OpenAI/Anthropic" a non-starter: regulated data, client confidentiality, source code you can't leak, or a data-residency requirement. Bionic keeps the agent on infrastructure you control — your laptop, your GPU, or a zero-retention cloud — instead of a third-party API that logs your prompts. If none of that constrains you, a hosted coding agent is probably still simpler; Bionic earns its place the moment privacy or cost is the deciding factor.
What it does#
- Code projects — point Bionic at a local folder and ask it to investigate, edit, or debug. Every change shows up as an inline diff you review as it goes, plus shell commands and repo-aware search.
- Work projects — hand it documents, PDFs, decks, and spreadsheets, or ask it to generate new ones. This is the "general knowledge work" surface, not just coding.
- Local voice input — a voice keyboard powered by Voxtral transcribes speech on-device, so dictation never leaves your machine.
Where the model runs — three modes#
This is the part that matters for a founder weighing privacy against horsepower:
- Local models — Bionic runs open models through the LM Studio runtime: GGUF (llama.cpp) everywhere, and MLX on Apple Silicon. Llama, DeepSeek, Qwen, Mistral, Gemma, Phi and hundreds more. No per-token cost beyond your own hardware.
- LM Link — route a job from Bionic on your laptop to another machine you own, like a home-lab GPU, over a Tailscale end-to-end-encrypted link. Heavy inference, still on your hardware.
- Secure Cloud — for models too large to run locally, LM Studio serves frontier open weights with Zero Data Retention and a commitment never to train on your data. Treat it like a managed open-model API — without provisioning an H100 cluster yourself.
Pricing and platforms#
The LM Studio desktop app is free and runs on macOS, Windows, and Linux. Local inference costs nothing beyond your own hardware and electricity. Secure Cloud is the paid, metered option for the largest open models — the escape hatch when local isn't enough.
How to start#
- Install LM Studio for your OS and open the Bionic app.
- Download an open model that fits your machine (an MLX build on Apple Silicon; a quantized GGUF elsewhere). If you're unsure which runner to use, our comparison of Ollama vs LM Studio vs Jan covers the trade-offs.
- Create a Code project, point it at a repo, and give it a small, bounded task — read this module, propose a fix — so you can watch the inline diffs before you trust it with more.
- Hit a model too big to run locally? Flip that job to Secure Cloud, or wire up LM Link to a GPU box you own.
Where it fits#
Bionic lands in the same lane as the growing "own your runtime" movement — the reason founders are asking where to serve an open model and where a long-running agent should live. The bet is the same one open weights have been making all year: as capable open models get cheap to run, the deciding question stops being "which model is smartest" and becomes "whose machine is it running on." Bionic's answer is yours.
