Mozilla’s Llamafile has always promised something close to absurdly convenient: an entire large-language-model runtime — weights and all — packed into a single executable that runs on almost anything with a CPU, with no installation step. Version 0.10.4, released this week, extends that “one file does everything” philosophy from text to speech, and quietly hardens the runtime in the process.
Llamafile works by combining llama.cpp with Cosmopolitan Libc into one framework that collapses the usual sprawl of an LLM setup down to a single-file binary. As the project’s own README explains, the goal is to make open models far more accessible to both developers and end users by removing the install-and-configure tax. The same packaging already produced whisperfile, a single-file speech-to-text tool built on whisper.cpp. With 0.10.4, that speech side gets a major upgrade.
The headliner is the first release of transcribefile. According to the 0.10.4 release notes on GitHub, transcribefile is a Cosmopolitan build of the command-line tool from transcribe.cpp — the ggml-based speech-to-text library from @cjpais that supports 16 or more model families. In practice, it means you can now grab a single executable that transcribes and translates audio locally, the same way llamafile already served up chat models. For a project whose entire thesis is “distribution should be one file,” folding speech transcription into that model is a natural step.
The speech capability sits on top of a younger project. Mozilla AI announced transcribe.cpp at the end of June as a C/C++ speech-to-text inference library positioned as a speech analogue to llama.cpp. It builds on GGML and can be accelerated with Vulkan, NVIDIA CUDA, and Apple Metal — the same accelerator backends that already make llamafile useful across wildly different hardware. The 0.10.4 release notes point readers to both the transcribe.cpp author’s post and Mozilla.ai’s announcement for deeper detail.
Image credit: Phoronix
Beyond the speech headline, 0.10.4 is a solid maintenance release with several fixes that matter for real-world use. The build pulls in a fresh sync with upstream llama.cpp (noted in the notes as a catch-up after too much time had passed between releases) and brings a handful of improvements to its Vulkan API and AMD ROCm acceleration handling. It also adds HTTPS support — both for the built-in server and for downloading models from Hugging Face by their IDs — which closes an obvious gap for anyone running the tool in anything resembling a production setting. Rounding out the release is the return of pledge/SECCOMP sandboxing, a security measure that restricts what the process can do after startup.
Why any of this matters comes down to the local-first philosophy Llamafile embodies. Running models on your own machine, with no round trip to a cloud API, keeps data private and removes a recurring bill — and the single-file format means a developer can hand a colleague or a user one executable and trust it will run. Extending that to speech-to-text widens the surface area: transcription, translation, and chat all become things you can ship as one portable artifact rather than a stack of dependencies.
The 0.10.4 changelog also lists a healthy stream of community pull requests — GPU device-count probing on Windows, documentation moves, ROCm script improvements, clearer Vulkan error messaging, and a fix for a Ctrl+C bug in chat mode — which is itself a signal worth noting. A release this broad across acceleration backends, networking, and security is the mark of a project with an active contributor base rather than a stalled experiment.
For users who already rely on llamafile for local chat models, 0.10.4 is a routine update that happens to open a new door: point the new transcribefile binary at an audio file and you get local transcription without standing up a separate service. For anyone evaluating self-hosted AI tooling, it is a tidy demonstration of how far the “one file, runs anywhere” idea has come — from text, to images, and now to speech.
