LM Studio Bionic (2026): what it is & which open models to run

On July 16, 2026, LM Studio shipped Bionic, its first standalone AI agent app for open models, on macOS and Windows. It links to a local folder for code, sandboxes your documents for work, and runs models on your own machine or routes heavy tasks to open models in the cloud. Here is what it actually does, how it differs from Claude Code, Cursor and Ollama, and which two open models it names, with what they cost in our executed benchmark.

What LM Studio Bionic is

LM Studio Bionic, announced on July 16, 2026, is described by its makers as the AI agent made for open models. It is the company's first move beyond its chat and runtime roots into a full agentic app. At launch it is available on macOS and Windows. The primary download page shows macOS darwin/arm64 builds, and coverage from 9to5Mac confirms Mac and Windows; Linux is not confirmed by these sources at launch.

Two ideas make Bionic worth understanding. First, it is an agent, not just a place to chat with a model: it plans, edits files, searches code and can roll work back. Second, it is built around open models, which you can run on your own machine or route to open models in LM Studio's cloud when a task exceeds your hardware. It also ships a cross-app voice keyboard whose dictation is transcribed locally, on-device, using Mistral's Voxtral speech model, so nothing leaves your machine for transcription.

Bionic is a separate app from LM Studio. You can keep using LM Studio alongside it for advanced low-level configuration while Bionic handles the agentic layer. If you are new to running models on your own hardware, our guide to running LLMs locally in 2026 covers the runtime basics that sit underneath an app like this.

Code Projects vs Work Projects

Bionic organizes work into two project types.

Code Projects. You point Bionic at a local folder and it can investigate, edit or debug the code. It supports agentic code search across the project, multi-file editing and code review, and it shows inline diffs so you can inspect every change before accepting it. This is the flow most people will compare to Claude Code or Cursor.

Work Projects. You load documents such as PDFs, decks and spreadsheets into a sandboxed environment. Bionic can organize local directories, edit and create files, and summarize materials. It includes native web search, automatic checkpoints so you can roll back, and in-app previews, with more file types planned. This is the part that pushes Bionic past a pure coding tool into general knowledge work.

Bionic vs Claude Code vs Ollama

The quickest way to place Bionic is against the two tools people already know. Claude Code is an agentic coding tool built on closed, hosted frontier models. Ollama is a bare model runtime that serves an API for other apps. Bionic sits between them: it is agentic like Claude Code, but open and local-first like Ollama, and it adds document work and voice. The comparison below is our analysis, since no source directly compares these three.

If you are shopping the runtime layer specifically, our roundup of Ollama alternatives is the companion read; Bionic is not really an Ollama competitor so much as an agent that can sit on top of a runtime like it.

Which open models to run

For Code Projects, LM Studio names two powerful open models it works with: GLM 5.2 and Kimi K2.7 Code. Note the wording: these are examples of open models it works with, not an official ranked leaderboard. Still, they are a useful starting point, and here is where DataLLM Lab can add something no launch coverage does.

We ran a first-party executed coding benchmark in July 2026: 13 models across 9 tasks where the model generates code and we then run hidden tests against it, with cost computed as token usage times list price per 1000 tasks. Both models Bionic names scored a perfect 9 of 9. GLM 5.2 landed at $1.99 per 1000 tasks at 12.3s; Kimi K2.7 Code at $1.34 per 1000 tasks at 10.4s, using just 272 reasoning tokens per task. In other words, the two models Bionic tells you to run are the same two our executed test rates 9/9 — and here is what they actually cost. Read the full method in our cheapest LLM for coding benchmark, and the model deep dive in our GLM 5.2 review.

Context matters: 10 of our 13 models scored 9/9, so single-shot correctness on standard tasks is table stakes. The real axis is cost, which spans about 88x across the field, from Qwen3 Coder Next at $0.10 to GPT-5.5 at $8.83. That is the chart below.

Local vs cloud, and what it costs

Bionic runs models via the LM Studio runtime locally, or switches to open frontier models in LM Studio Secure Cloud for heavier tasks that exceed your hardware. Cloud requests run under a Zero Data Retention policy by default: they are processed transiently and not retained after the request completes, and LM Studio commits to never training on user data. Cloud access requires an LM Studio account with billing set up.

A simple decision rule: if the model and context fit comfortably on your machine, run local, because it is free, private and offline. If a job is too big or too slow locally, route to the cloud. Just note the pricing caveat: no launch source published a Bionic subscription or a cloud rate card, so the $1.99 and $1.34 figures above are our list-price cost per 1000 tasks, not LM Studio's cloud billing. Use them to compare models, not to predict your LM Studio invoice.

One more launch-day data point on capacity. We ran Kimi K3 on launch day, July 16, 2026, via OpenRouter: it passed all 8 tasks we could execute across two runs. The 9th never completed due to repeated upstream 429 rate-limiting — a launch-day capacity signal, not a capability miss. Its real cost landed around $4 per 1000 tasks, in Claude Opus 4.8 territory, while staying token-lean at roughly 150 reasoning tokens per task. See live pricing on our pricing page.

The cloud path and a gateway

Bionic's cloud path is deliberately open-model only, which is the whole point of the app. That also means you can serve those same open models yourself. If you want the cloud path but on your own terms — one API key, usage you control, and the ability to swap between GLM 5.2, Kimi K2.7 Code and others — an OpenAI-compatible gateway does the job. It is the same pattern people use to point coding agents at any model; our writeup on a Claude Code router shows the routing idea in a different tool.

The unique claim DataLLM Lab can make is narrow and testable: the two models Bionic tells you to run are the two our executed benchmark rates 9/9, we have published what they cost, and our gateway lets the cloud path reach either one. Start local, lean on the cloud only when a task demands it, and you get the best of both without guessing.

FAQ