llamafile update 20260202

[aittalam]

Hello everyone, here’s our latest update for llamafile!

Since our Christmas update that brought in the TUI and metal support in our WIP branch, we updated llama.cpp to f47edb8 (5 January 2026) and implemented several key improvements:

• Added zipalign as a GitHub submodule (so we can get the latest updates from Justine’s repo)
• Brought back cuda support on Linux
• Added support for the mtmd API in the TUI (so you can now directly access modern multimodal models from the llamafile chat)
• Tested new llamafiles running models trained for tool calling (e.g. Qwen3, gpt-oss-20b) and multimodal models such as llava 1.6, Qwen3-VL and Ministral 3

Overall this has been a great learning experience. We promised we would have shared our learnings too, so here’s more information for you. First of all, our process:

• if you follow our merged PRs history chronologically, you’ll see the impact of adding individual new features to our baseline (the minimal cosmocc build of llama.cpp)
• our latest list of patches shows the actual changes applied to the llama.cpp submodule

I personally also collected a lot of notes about llamafile: some of them started from an analysis of the old code and of the diffs wrt the old llama.cpp, others from bugs investigations, most have been at least partially generated by Claude (more info about this below). Here are a few I think could be interesting, let us know if you find them useful:

• Fixing Cross-Module Memory Management in Llamafile's Metal Backend
• CPU Optimized Kernels Workstream Summary

Last but not least, as I talked about using AI, I would like this to be as open as the code itself. I have found Claude useful not just for code analysis/writing or wrapping up my notes and code udpates into documents, but also in automating a set of tasks that I would otherwise repeat manually (e.g. “test this bugfix by rebuilding the code, running llamafile with these parameters and checking if and where it segfaults” 😅). In the spirit of transparency, we are sharing:

• Markdown files describing the llamafile project, the build system, and instructions on how to rebuild, patch, test, and keep the project updated with upstream llama.cpp. While we are using these as “agent skills” to automate menial tasks, they are human readable and can be useful to all contributors, whether they write code manually or using any AI-assisted tool
• AI-generated PRs (for less critical tasks such as update-llama-cpp), so you can follow both the automated steps and our manual iterations, and possibly contribute not just to the code but to the process too.

All the updates are in our WIP branch, and more recent alpha binaries are available here.