Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Inference & Quantization / Audio & Speech / Inference & Hardware / Models & Weights

llama.cpp 0.9.0 expands local model and audio support

The project’s November release adds FP8 support, multi-sequence cache work, additional architectures and audio features.

Listen to this articleBrowser Voice
Ready
Concept illustration of a local AI workstation
AI-generated concept illustration created for OpenSourceCoder.com. It is not an official product screenshot or a photograph of an actual event.
Event / context: llama.cpp release: November 4, 2026. Source check: Nov 5, 2026.

The llama.cpp 0.9.0 release on November 4 adds an FP8 E4M3 tensor type, work on unified cache handling for multiple sequences, and additional supported model architectures.

Why it matters

The maintainers also describe expanded speech support, including Chatterbox text-to-speech and Nemotron 3.5 transcription, alongside updates to ggml and fixes to the server and interface. A feature appearing in release notes does not mean every backend, model file or machine supports it equally.

For people running models at home, check the actual build, hardware backend, available memory and separate model licenses before switching. We have not benchmarked this version, and these release notes do not establish a speed improvement on our RTX 3060 setup.

Explore the llama.cpp project profile or read the linked upstream notes for the complete changes.

EXPLORE THE SOURCE

GitHub repositories

Reported from linked project sources. OpenSourceCoder.com has not independently installed or tested this release.

Follow your curiosity.

Pass the signal.