LM Studio Bionic: What It Is, Pricing, Privacy and How It Differs From LM Studio

LM Studio Bionic is a separate desktop agent app from the makers of LM Studio, released on 16 July 2026. It researches, drafts documents and slides, and edits code with a model on your computer, a model on another of your devices through LM Link, or a US-hosted open model in LM Studio Secure Cloud. Local use is free; hosted-model plans start at $20 a month.

Questions about Bionic circle four points: whether it replaces LM Studio, what it costs, where your prompts go, and how it compares with coding agents such as OpenCode. The answers below come from LM Studio’s documentation, changelog, pricing, terms and privacy policy, and from the other projects’ repositories and docs, checked 14 October 2026, when the newest release was Bionic 1.1.8 (12 October).

If all you need is to run models and expose them to other software, classic LM Studio is still the tool, and our inference server comparison weighs it against Ollama, llama.cpp and vLLM.

What LM Studio Bionic is

LM Studio’s documentation calls Bionic a new, separate app, built for agentic work with open models. Work happens in projects, which share files and context, and in sessions, the task threads inside a project. Sessions keep running in the background, can be forked into parallel branches, and can split a large coding task into sub-sessions.

The launch version had two project types, Work Projects and Code Projects. Version 1.1.0, released on 27 August 2026, merged them into one project type with an optional Allow coding switch:

• Documents and research. Bionic reads, edits and creates documents, PDFs, slides, spreadsheets and images in a managed sandbox, with in-app previews (Office and HTML previews arrived in 1.1.8) and automatic checkpoints you can roll back.
• Code. With coding allowed, you choose a working directory. Bionic indexes it and can search the repository, edit files, use Git and run shell commands inside that folder, showing each edit as an inline diff.

Voice input runs Mistral AI’s Voxtral transcription model on the device, and the launch post says the voice keyboard can dictate into any app. Skills use the Agent Skills format (SKILL.md), and Bionic can reuse skills you keep for Claude Code or Codex.

Where the model runs

LM Studio Bionic vs LM Studio

The two apps share a runtime but are built for different jobs.

If you are looking for Bionic’s developer mode, it lives in the other app. Developer Mode is a classic LM Studio setting under Settings > Developer, and Bionic’s documentation has no equivalent; it suggests using LM Studio alongside Bionic for advanced low-level configuration.

Since 12 October the two also cooperate directly. With Bionic 1.1.8 and LM Studio 0.4.26 or newer, both can run at once, and LM Studio’s changelog says Bionic then uses LM Studio for local model inference. That lets you keep LM Studio’s local server available to other tools while you work in Bionic.

If you are choosing a desktop model runner rather than an agent, our Jan and LM Studio comparison covers the open-source alternative, and Unsloth Studio against LM Studio covers fine-tuning.

Pricing, terms and licence

Credits pay for usage beyond a plan’s allowance, and Bionic reports consumption over the last 30 days rather than a price per task. For groups, LM Studio offers organisations with centralised billing for inference credits and says team subscription plans are coming soon. Its enterprise page takes sales enquiries from companies that want Bionic on LM Studio’s cloud, their own infrastructure or edge devices.

For work use, the terms matter more than the price:

• Source. Bionic is proprietary. LM Studio’s GitHub organisation publishes its CLI, SDKs and MLX engine under MIT, but no Bionic code.
• App terms. The desktop app terms (version of 23 August 2026) license the software for personal and internal business purposes only, and exclude service-bureau and software-as-a-service use, modification and reverse engineering. They do not name Bionic, so confirm with LM Studio which terms cover a company rollout.
• Paid services. LM Studio’s terms of the same date say fees are non-refundable unless the law or the offer says otherwise, unused allowances do not roll over, and usage credits may expire.

LM Studio Bionic download and setup

LM Studio’s download server offered Bionic 1.1.8 for these systems on 14 October 2026:

• macOS on Apple silicon (.dmg); there was no Intel Mac build.
• Windows on x64 and ARM64 (.exe).
• Linux on x64 and ARM64 (AppImage), supported since version 1.1.2 on 8 September.

LM Studio publishes no separate requirements page for Bionic. The classic app’s page is a fair proxy: macOS 14 or newer with 16 GB of RAM or more recommended, AVX2 on x64 Windows PCs, Ubuntu 20.04 or newer on Linux, and no support for Intel Macs.

1. Download the installer for your system from LM Studio’s own download page rather than a third-party mirror. On Linux, make the AppImage executable before the first start:

   chmod +x ./Bionic-*.AppImage
   ./Bionic-*.AppImage

2. Skip signing in if you will only use models on your own machine; hosted models, credits and LM Link need an account.

3. Open Settings > Local Models > Explore, filter by format, and use the device-fit information to choose a variant that fits. Our VRAM calculator shows how much context a model leaves on a given GPU or Mac, and the Mac memory guide lists picks by memory tier.

4. Under Settings > General, choose a local model as the Root model, so every new session starts on a model you picked.

5. Create a project. For a codebase, switch on Allow coding and select the repository’s root folder.

6. Before the first coding task, pick the shell mode in the composer: manual review or Auto Review, not allow-all.

7. If you also use LM Studio, update it to 0.4.26 or newer so the two apps can run side by side.

For coding, reliable tool calling matters more than parameter count; our shortlist of local coding models covers the candidates.

Privacy and safety settings

Bionic’s privacy is decided session by session, by the model and the tools you allow.

Hosted models and web search

A session on a Secure Cloud model sends its prompts to LM Studio’s US-hosted service. The launch post commits to zero data retention and no training on user data for all Bionic users. LM Studio’s app privacy policy, effective June 2026, adds that cloud requests are processed transiently, are not kept after completion by LM Studio or its providers, and are transferred to the United States. European teams should raise that transfer with their data protection officer; our zero data retention explainer covers what such promises do and do not include.

Web search has its own switch under Settings > General > Web Search. LM Studio says searches also run under zero data retention but warns that web content should be treated as untrusted. Keep it off in projects that hold confidential files.

Shell modes and Auto Review

The composer offers four shell modes: disabled, manual review, Auto Review and allow-all. Auto Review, added in 1.1.0, first parses each command and approves only patterns it recognises as safe, without asking a model. Commands it cannot clear go to a reviewer sub-agent that rates the risk and whether you asked for the action. Commands rated too destructive never run, and rejections come back to you rather than to the model. LM Studio’s own write-up concedes that complete protection against prompt injection is impossible, so use manual review in repositories that hold credentials or deployment scripts.

Skills and external files

Skills install from a URL, with @Install Skill, or as a SKILL.md file under Settings > Skills, where skills found in other apps can be switched off. A skill is a set of instructions the agent will follow, so read third-party skills first, as LM Studio’s docs also advise. Files added as external files stay where they are and are edited in place; give Bionic copies when the source matters.

LM Studio Bionic vs OpenCode, Hermes Agent and other agents

OpenCode and Hermes Agent come up in the same searches, but they are open-source agent harnesses that you connect to a model of your choice, while Bionic ships its own runtime and hosted models.

Bionic vs OpenCode. Choose Bionic for one graphical app that downloads and runs models and also handles documents and slides. Choose OpenCode if you work in the terminal, want an MIT licence or need a provider Bionic does not offer. They are not exclusive: OpenCode can call LM Studio’s server at http://127.0.0.1:1234/v1, as our OpenCode setup guide shows.

Bionic vs Hermes Agent. Hermes lives on a machine you control, reaches you through chat apps and requires a model context of at least 64,000 tokens. Bionic stays on your desktop, built around files you can see and roll back.

Other names that come up next to Bionic:

• Claude Code is Anthropic’s coding agent. LM Studio’s January 2026 guide points it at LM Studio’s Anthropic-compatible endpoint (version 0.4.1 or newer); our Claude Code setup guide has the steps.
• OpenClaw is an MIT-licensed personal assistant, stewarded by the OpenClaw Foundation, that answers in more than 20 chat channels.
• DeepSeek Harness (dsh) is DeepSeek’s MIT-licensed harness, published on GitHub on 13 August 2026 as a developer preview with a web UI.
• Ollama runs models rather than agents, so it competes with classic LM Studio, not with Bionic.

For a deeper look at the open-source options, see our Hermes Agent, Claude Code and OpenCode comparison.

LM Studio Bionic review: strengths, limits and a work checklist

This assessment rests on LM Studio’s documentation, changelog and pricing as of 14 October 2026; we have not benchmarked speed or output quality.

Strengths

• One app covers local, LAN-linked and hosted open models with the same projects and sessions.
• Documents are a first-class use: Office previews, a sandboxed workspace and checkpoints.
• Shell safety is documented, with a deterministic first pass before any model is consulted.
• Local use is free, voice included, and releases are frequent: nine versions from 1.1.0 on 27 August to 1.1.8 on 12 October.

Limits

• Closed source, so a security review depends on what LM Studio chooses to document.
• Hosted models run only in the United States.
• Team subscriptions are listed as coming soon, and the docs and pricing page describe no single sign-on or central policy controls beyond organisation-managed MCP servers.
• No published system requirements for Bionic, and no Intel Mac build.

Before using it at work

1. Check that your AI policy covers desktop agents staff install themselves; our shadow AI explainer explains the gap.
2. Decide which projects may use hosted models or web search, and set a local Root model.
3. Confirm the terms fit: internal business use is allowed, offering Bionic to others as a service is not.
4. Set the shell mode on every coding project and keep secrets out of working directories.
5. In the EU or UK, check that sending hosted-model prompts to the United States is covered by your data protection assessment.

If several people need one shared chat interface rather than individual installs, start with our comparison of self-hosted chat interfaces.

How VDF AI fits

Bionic is designed for one person on one computer, and its safeguards are personal ones: a shell mode you choose, a document sandbox and a retention promise for its cloud. A company putting agents in front of hundreds of employees has to answer who may use which model and data, what each agent did, and whether any prompt left the building.

VDF AI is the governed on-premises alternative for that case. It runs on your own servers, or as VDF AI Cloud, and provides governed agents and workflows, private RAG over company data, chat and VDF Code for engineering teams, with role-based access control on every plan and audit logs. On-premises deployments sign staff in through built-in Microsoft Entra ID single sign-on. VDF AI does not run, manage or connect to Bionic or LM Studio.

At the model layer, VDF AI Router gives applications a single endpoint. An Ollama server or another on-premises model deployment can sit behind it next to the cloud models your policy approves, and air-gap mode confines routing to local models.