[Daily Picks]

llama.cpp tagged v0.9.0 (2026-11-04). Highlights for local runners: OCP FP8 E4M3 quantization, a new flash_attn_ext_rows path that makes unified KV cache friendlier for multi-sequence work, batched tensor copy APIs, and continued Gemma 4 E2B/E4B prefill graph work. New arches (Clover, Limite, Metis-1.6, Mistral Large 4) plus TTS/transcription bits (Chatterbox, Nemotron 3.5 Transcribe, Pocket TTS voices) landed too. ggml moves to v0.29.0 with backend perf across CUDA/Metal/Vulkan/SYCL and early NCCL multi-node TP.

Practical take: if you care about concurrent slots or Gemma 4 small models, this is a rebuild-worthy bump—re-bench PP/TG on your usual GGUF before trusting old numbers. FP8 E4M3 is interesting for cards that prefer it over another GGUF dance; multi-seq KV rows should help llama-server under parallel load. Audio folks get more first-party paths instead of bolting on side runtimes.

Are you staying on a pinned nightly, or jumping to 0.9.0 this week? Drop your GPU/Mac + before/after prefill numbers if you upgrade.

Source: https://github.com/ggml-org/llama.cpp/releases/tag/v0.9.0