LM Studio Bionic. How local is your agent?

As of 4 November 2026. You have two supplier quotations and want a comparison spreadsheet. LM Studio Bionic can read the PDFs and create a file. Whether their contents stay on your computer depends on where the model runs and which tools it may invoke while reading and writing.

Three places to run inference

With a remote model, the vendor says the conversation stays on the device you are using. The information required for inference is processed by the other computer. For cloud models, LM Studio describes transient processing with zero retention. That is the vendor’s commitment. Its billing documentation says local and LM Link inference consume no LM Studio credits. Model location and data flow · Credit accounting

What changed in October

Bionic 1.1.10, released on 29 October, brings the in-app terminal to public builds. If a project needs a script, you can inspect the command and its effect in the same workflow. The release also changes the presentation of code diffs and adopts a more conservative voice default. On Macs with less than 32 GB of memory, the voice model no longer loads automatically by default. Version 1.1.8 had already added previews for Word documents, spreadsheets and presentations, and support for running alongside classic LM Studio 0.4.26 or later. 1.1.10 release notes · Version history

A model’s size on disk is only a starting point when buying hardware. Available memory must also cover the active context and other programs. A universal promise that 32 GB is enough for coding would ignore both the model and the job. Start with a smaller local model, check its tool support and load a representative document set before ordering hardware. Choosing local models

A useful test with two quotations

1. Put two fictitious quotes in a dedicated project folder. Change the minimum order in the second one and put a delivery charge in a footnote. Ask for unit price, minimum order, delivery and a source reference for each value. Missing values should stay blank.
2. Select a local model explicitly. Leave web search and external services off for this task. For confidential work, also check whether a separate model handles images or another step. Voice input can transcribe locally. The resulting text is then used by the session’s chosen model. Document tools · Voice input
3. Open the result independently and calculate one order value by hand. If a value is wrong, inspect the passage cited in the PDF. A missing source, a misread number and a broken formula call for different fixes.
4. For a second run, change one setting. Record elapsed time, peak memory use and errors. This helps distinguish a model limitation from document extraction or spreadsheet logic.

In a code project, Bionic can search, edit files and execute shell commands. A working copy with a clean Git status makes a good first trial because you can inspect the diff line by line. This assessment draws on product documentation. Actual timing depends on the computer, model and workload. Coding documentation · Local AI and cloud compared