On 16 July 2026, LM Studio announced Bionic, which it describes as an AI agent built for open models. The first thing to understand is structural rather than technical: Bionic is a new, separate application. It does not replace LM Studio. LM Studio continues to exist, and the announcement points people who want low-level configuration back to it.
That framing matters because the two products answer different questions. LM Studio asks what you want to run and how. Bionic asks what you want done and lets you choose where to run it.
What Bionic is
Bionic is pitched at real work: coding, research, and tasks that involve documents and files. It brings together four things.
An agent for coding. You create a Code project and point it at a local folder. Bionic can inspect the codebase, explain unfamiliar code, trace behaviour through agentic code search, and make edits, with inline diffs for reviewing each change.
An agent for documents. In a Work project, Bionic handles documents, PDFs, decks, and spreadsheets, and can generate new ones. It processes documents in a sandboxed environment, which LM Studio says keeps the rest of your machine and files out of reach. It offers automatic checkpoints for rolling back changes, in-app previews, and native web search for pulling in outside context.
Local voice input. Bionic ships a voice keyboard with on-device transcription that works from any application, inserting text wherever your cursor is. At launch the transcription model is Voxtral, from Mistral AI. LM Studio states this runs entirely locally.
Flexible model execution. This is the crux, and it gets its own section.
The three execution paths
Bionic gives you three places a model can run, and the privacy answer is different for each.
Local. You download models inside Bionic and run inference on your own hardware. The selected model processes the prompt locally; web search, remote tools, and other connections still need separate attention.
LM Link. A connection path to models running elsewhere that you control.
LM Studio Secure Cloud. Frontier open models, served by LM Studio, for tasks the local machine cannot carry. Using cloud models requires creating an LM Studio account and setting up billing.
The announcement is candid that the point of the design is choosing per task, and it frames this partly as cost control: pick the right model and the right compute environment for each job. The named cloud-scale models in the coding section are GLM 5.2 and Kimi K2.7 Code, neither of which is a laptop proposition.
So the answer to "is Bionic local" is that Bionic is local-capable and hybrid by design. It is not a local-only tool, and it does not claim to be.
Reading the zero-retention commitment carefully
LM Studio states that for all Bionic users it commits to zero data retention and to never training on user data, and that cloud requests are processed transiently and are not retained once the request completes.
This is LM Studio's stated cloud-processing commitment. Local inference and a hosted service remain different data paths; inspect the applicable contract, policy, and settings for the selected path.
If your threat model includes a compliance obligation, client confidentiality, or data-residency requirement, the local path answers a different question from the cloud path. The cloud path requires review of the applicable contract, policy, and technical controls.
Which network paths exist
For readers auditing egress, these are the features that leave the device by design:
- Cloud model inference through LM Studio Secure Cloud, plus the account and billing relationship it requires.
- Native web search in Work projects, which fetches outside content into your workflow.
- LM Link, depending on what it is linked to.
And the ones that do not, per LM Studio:
- Local model inference through the LM Studio runtime.
- Voice transcription, which runs on-device using the bundled audio model.
The sandbox around Work project document processing is worth naming separately, because it is a different kind of protection. It limits what Bionic can reach on your filesystem. It is not a network control, and it is not a substitute for one.
Should you switch?
Bionic and LM Studio can sit side by side. LM Studio remains a workspace for advanced model configuration, but Bionic also has runtime settings: Settings > Runtime now offers Splash under Experimental backends on compatible Macs. Runtime selection is therefore not exclusive to the original app. Follow the Qwen3.8 Splash setup for the exact Mac requirements and model package.
Reach for Bionic if you want an agent that will actually work through a codebase or a pile of documents, and you are comfortable choosing per task whether that work happens locally or in a hosted environment.
Stay with LM Studio if what you valued was a local inference application with a settings panel and no account, and you would rather not have a billing relationship in the loop at all.
Run both if you want the agent for the heavy tasks and the older app for the tuning. That appears to be the arrangement LM Studio expects.
One practical note before you download
The current download page offers Bionic for Windows as well as a separate download for the original LM Studio app. Choose the application and operating system deliberately. Bionic availability on Windows does not make Splash a Windows engine: the published Splash integration requires an M3-or-newer Mac, macOS 26.4 or later, and at least 36GB of unified memory. Confirm a Linux build on the download page before planning around it.