0.3.3

localm 0.3.3 adds Kimi-Linear support and Spanish and French, and fixes a long list of everyday problems in updates, model switching, memory and the coder. The full list is in the changelog under 0.3.3.

New

• Kimi-Linear models work out of the box, from the safetensors repository as well as a GGUF. A checkpoint larger than your memory keeps its routed experts on disk and reads them as needed, so the 48B model answers on a 16 GB card, slowly; the load output says so.
• A model that ships its own code runs without it when transformers supports it, and the Models page and load output say so.
• The interface can be shown in Spanish and French.
• Every model on the Models page can be renamed, given an alias and removed, and a rename updates what named it. The page also has a search box and a reload button.
• "Remember that ..." in a chat message is saved to memory.
• localm can add the packages a trusted model's own code needs, after showing you the list.
• Settings > Runtime & GPU has Unload other models to make room and Context window (minimum).
• The ExLlama plugin can be installed from setup and the Plugins page (NVIDIA, Linux).

Updates

• An update from the app installs the new native runtime and components without a manual restart, and no longer rolls back because the app itself is using the runtime.
• Updates and rollbacks move every part of the install together. Settings > Updates says why an update was undone and what to do.
• Setup names any localm programs running from its folder and asks before stopping them, so running it with localm still open no longer fails halfway.

Fixed

• Long prompts no longer fail on gpt-oss, Llama 4, Mistral Nemo and other models whose tokenizer could not handle very long text.
• After you switch models, your next message is answered by the model you switched to.
• Agentic coding no longer fails with "exceeds the model's maximum context capacity".
• Deleting a chat from the sidebar asks first.
• A crafted model or document file can no longer crash, freeze or swell localm while it is read.
• Generating with a missing image, video or music model offers the usual download box.

Runtimes

• llama.cpp moved to b12087 and the AMD ROCm runtime to b1363. An update installs them on its own.

Updating from 0.3.2 on Windows

• If localm was started with localm.exe, localm.bat or the global localm command, the update leaves that one launcher file in place and carries on. If an update started that way still rolls back, close localm and update from the desktop shortcut, or download this release and run setup again.
• Clients from 0.3.1 and earlier are offered 0.3.2 first, then this release.

The next release (0.3.4) will be mostly bug fixes. If something's broken for you, or there's a model or feature you'd like added after that, feel free to open an issue or start a discussion.