0.3.2

localm 0.3.2 makes updates complete, brings the multi-GPU overhaul, and fixes a set of everyday problems in the coder, model switching and setup. The full list is in the changelog under 0.3.2, 0.3.2rc2 and 0.3.2rc1.

Updates bring everything up to date

• Updating localm now also updates the components it installed: the llama.cpp runtime for your backend (the AMD ROCm build included), PyTorch and transformers, the image (stable-diffusion.cpp) and music (KoboldCpp) runtimes, localm's managed ComfyUI, the uv inside localm, and the CUDA libraries on Linux. Before, an updated install kept the versions it was first set up with.
• A component you never installed is never installed, and one you chose yourself is kept and named in the update summary.
• localm stops its own processes that use a component before replacing it. If anything fails, everything is put back and you are told what failed. localm update --rollback restores the previous versions too.
• An install updated by an earlier version catches up the next time you start a chat, localm serve, the GUI or a media command.
• localm setup-llama no longer removes a working runtime when the new one cannot be downloaded.

Models and GPUs

• Switching models or pressing Unload while a reply is being written asks first instead of cutting the reply off. Cancel keeps the model and lets the reply finish. API clients can confirm with force.
• Picking a model that does not fit even after unloading the current one now asks before unloading anything; Cancel keeps your chat model and the embedding model loaded.
• Unload all also stops localm's own idle image, video and music workers, and loading a chat model frees them first, so the chat model is not pushed onto the CPU by localm's own image worker.
• On an AMD GPU under Windows without PyTorch, automatic GPU layers now count the VRAM other programs use, so a model no longer spills into slow shared memory.
• The GPU numbers you pick in Settings are the ones the GPU selector shows, and ticked GPUs are filled in the order you ticked them.
• A model set to 0 GPU layers now runs entirely on the CPU. Before, an AMD GPU could still be used for part of the work.
• An image, video or track made with a model picked on its page now frees the VRAM that model needs, not the amount the Settings model needs.
• When localm's own ComfyUI stops working after a GPU error, the next image or video job restarts it instead of failing or starting a second copy.
• Wan video on AMD GPUs no longer fails at the last step with a GPU error or a driver reset: localm's ComfyUI now decodes the video in tiles there.

Coder

• --max-turns limits each task again, not the whole session.
• No more hundreds of llama.cpp loader lines in the session, no false "saved session could not be read" warning, and a recovered GPU embedding error is one warning line instead of a traceback.
• The task list the model writes is shown as a checklist.
• A request to read or summarise without editing ("do not edit anything") finishes normally instead of being pushed to write a file.
• localm coder doctor -c <project> starts the project's MCP servers in the project folder.
• A scheduled coder job in a project set to privacy mode no longer leaves a session log.

GUI and setup

• Settings > Updates keeps the update options next to Check for updates, and shows release notes as formatted text.
• Error messages stay until you close them; Settings > Recent errors lists the last five.
• Settings no longer shows "Could not check the browser." when the browser plugin is not installed, and a failed load from the Models page no longer leaves the status on "loading".
• An audio or video clip sent with a short message gets a full player, and parameters set for one chat no longer carry into new chats.
• Every setup offers the same choices, and installing uv during setup no longer leaves its download in your system temp folder.

Privacy

• A key without the owner scope no longer sees folders on the host in bug reports, the model list or error messages; the owner still sees full paths.

Getting it

• From 0.3.1 or a 0.3.2 release candidate: update from Settings > Updates as usual.
• After updating from inside localm, close localm completely and start it again once, so it can bring the llama.cpp runtime and the other components up to date. A fix for this follows in the next release.
• From 0.3.0: the update also reinstalls dependencies, which needs a uv command the app can find. If uv is not installed on your system, the update stops and leaves 0.3.0 as it was. Then either install uv (for example with Astral's installer), close localm, start it again from a new terminal window (or after signing out and back in) so it sees the updated PATH, and update again; or extract this release over your install folder and run setup.bat (Windows) or bash setup.sh (Linux, macOS). Your models, chats and settings are kept.
• A fresh install: download the zip and run setup.bat (Windows) or setup.sh (Linux, macOS).