llama.cpp v0.9.0 ships FP8 types and expanded audio interfaces, not universal speedups
Version 0.9.0 was released within this window, grouping FP8 E4M3, multimodal and backend updates. The release is new, but not every included feature was merged today. Format support, backend availability and workload gains remain separate checks.
Source: ggml-org/llama.cpp 官方发布
llama.cpp released v0.9.0 at 13:08 UTC on November 4. Its release page identifies b12110 as the corresponding nightly build. This is a new version milestone, not evidence that every included change first appeared today. For local deployment, the boundaries of formats and interfaces matter more than treating release notes as a speed promise.
-
FP8 support refers to a specific representation. The notes list OCP FP8 E4M3 in saturating mode, conversion helpers and a new model file type. The linked PR explicitly separates KV cache, CUDA and Vulkan work into other changes. Recognizing a format does not establish a complete, efficient path on every backend. FP8, NVFP4 and MXFP4 must not be treated as interchangeable precisions.
-
Audio interfaces have concrete additions. The release includes Chatterbox TTS, Nemotron 3.5 Transcribe and predefined Pocket TTS voices, alongside related mtmd APIs. GET /props can report a model’s predefined voices and the TTS endpoint accepts their names. This confirms interfaces in the release, not a universal invocation scheme for audio models or permission to use all model weights commercially.
-
Pin the environment when upgrading. The notes also list LFM3 NVFP4 GGUF support, tensor splitting and backend optimizations, but establish no universal hardware gain. AraoAI has not independently benchmarked this build. Older PR tables are not fresh measurements, and prefill improvements do not imply proportional gains in token generation.
Keep the previous binary, arguments and model identity. Check loading, output correctness, time to first token and generation speed with fixed prompts and output lengths. Audio users should also check voice enumeration and input/output contracts. A version identifies the implementation; it does not prove that a configuration works.