llama.cppProduct update

llama.cpp · v0.3.0

llama.cpp 0.3.0 supports the dots3-note multimodal model with a new DSA-ISWA KV cache type, and GLM-4.5-Air gains MTP support. DeepSeek 4 gets fixes for tensor split (-sm tensor) and multi-sequence rollback. The bundled ggml library moves to 0.22.0, and mtmd adds dots3-note vision and audio along with WebP decoding. The server adds a LLAMA_SERVER_SLOTS_N_DIFF debug option, and the web UI gets tabbed chat navigation.

GitHub ReleasesAug 25, 20264 min read