For offline manuscript chat or a local API, I would start with the LM Studio desktop client. Choose Bionic when you need an agent that works with files and tools. Bionic can also use local models, but it offers cloud and remote routes too. Decide where requests will run before choosing a plan or assuming a session is offline.
We checked the official website’s interactive Bionic demonstration and pricing controls on October 28, 2026. This is a setup and privacy-boundary review, not an installed-app benchmark: we did not download software or models, generate a reply or audit network traffic.
The homepage now promotes Bionic; the desktop client still has its own section
The current homepage presents Bionic as an agent for work and code. Its interactive demo has a chat, a file preview and a model picker. We opened that picker: Kimi K3 was selected, and Qwen3 8B was another listed choice. We did not change the model or send a message.
That demonstration is useful for understanding the agent interface. It does not establish which models your computer can run, whether a real request stays local or how quickly a model responds.
The official download page separately lists LM Studio 0.4.26, described as a chat interface with a programmable API, below Bionic. For a writer who mainly wants to draft with a local model, that distinction matters more than the homepage’s agent features. You do not need to choose an agent workflow just because it is the main promotional entry.
Local, Remote and Cloud describe different inference routes
The Bionic model guide distinguishes these routes. This is the documented behavior; we did not exercise the connections:
- Local
- Your current computer runs the model. Model files must be available and fit the device's resources.
- Remote
- Another linked device runs the model. LM Link needs that device to be available and sharing its models.
- Cloud
- LM Studio Secure Cloud processes the request. The guide requires internet, a signed-in account, billing setup and available credits.
A remote model may be useful when a laptop lacks memory but another machine can handle inference. It is still a connection to another device, so it is not a substitute for a disconnected-laptop workflow. Do not confuse the name of the model with the location doing the work.
A local model does not make every agent tool offline
The LM Studio offline guide says downloaded-model chat, document processing and the local server can work without internet. Model searches, new downloads, catalog details and runtime updates require connectivity.
Prepare the model and runtime before disconnecting. A model name visible in the catalog is not evidence that the required files are already available. If catalog search fails offline, try the model you already prepared before concluding that local inference itself is broken.
Bionic adds another distinction: selecting a local model describes inference, while web search is an online operation. Our practical recommendation is to leave web search and other networked tools out of an offline trial. We have not verified a current toggle path, so this is a test boundary rather than a claim about a particular settings button.
On the public pricing page, we expanded the web-search ZDR question. The answer claims Zero Data Retention for search, while warning about sensitive queries and untrusted web material. No retention does not mean no transmission. The app privacy policy separately describes transient cloud processing when cloud services are used; that is a provider statement, not a privacy audit we performed.
For a private manuscript, decide separately whether you permit cloud inference, online research and remote-device processing. A blanket “local AI” label cannot answer all three questions.
Free local use and Bionic+ are different purchasing decisions
When checked October 28, the public price page listed a $0 Free option with local models and Bionic+ at USD $20 per month with hosted-model access. These are published monthly terms, not checkout prices we tested.
We also opened What are cloud credits for? Its answer says they cover use beyond a plan’s included allowance. Do not turn the subscription price into an assumption of unlimited requests.
The credits guide says local models and remote LM Link models do not consume cloud credits or require billing setup. Cloud consumption depends on the model and tokens. It documents Settings → Billing and Usage, a 30-day usage view and Refresh; we did not open that screen in an account.
If you only need on-device drafting, prove that workflow before buying cloud access. If you need hosted inference, check the actual allowance and model-specific rates in your account before budgeting a long writing session. We have not measured a cost per chapter or verified an exact included token quota.
Check both app and build when connecting Bionic to LM Studio
The LM Studio changelog lists 0.4.26 with an October 13 heading. Its Build 3 entry says both apps can run together: Bionic uses LM Studio for local inference while LM Studio is running, with Bionic 1.1.8 or newer required.
Later entries under the same version report Windows-on-ARM CUDA detection and lms compatibility fixes in Build 5. A version number alone can hide a relevant build difference. If that setup fails, record both apps’ versions and the LM Studio build before changing models or reinstalling anything. We did not reproduce those bugs or validate the fixes.
Check hardware as well. The desktop requirements require AVX2 on Windows x64 and recommend 16GB RAM; Apple Silicon Macs require macOS 14 or newer, with Intel Macs currently unsupported. The RAM recommendation is not a promise that any particular model or context length will fit.
Next: prove one small offline writing session
Choose the desktop client for chat/API work, or Bionic with a deliberately selected local route for an agent workflow. Then run this proposed check with harmless fictional material:
Before trusting the setup: a three-part offline check
- Prepare while online. Record the app/build, model, runtime and context setting. Confirm the files are available before disconnecting.
- Test the actual workflow offline. Start a short fictional writing session without web search or networked tools. If you need document chat or another frontend's local API connection, check that path separately.
- Keep the result narrow. A successful reply establishes that tested configuration's operation. It does not prove model quality, every tool's data flow or server-side retention.
We have not run that offline test. For a different local runtime and API workflow, compare the existing Ollama review; for a story-focused local interface, read the KoboldCpp review. Pick the workflow first, then the model and any optional hosted service.