Tools
llama.cpp 0.3.0 adds the dots3-note multimodal model with a new DSA-ISWA KV cache, plus GLM-4.5-Air multi-token prediction
Published 2026-08-25 at 10:22 UTC, llama.cpp v0.3.0 introduces dots3-note with a DSA-ISWA KV cache type (PR #27060), MTP support for GLM-4.5-Air (#26534), and DeepSeek 4 tensor-split via `-sm tensor` (#26490) alongside a multi-sequence rollback fix (#26756). ggml moves to v0.22.0 with meta-backend tensor split and per-op Metal kernels compiled in parallel, and mtmd picks up WebP decoding and a Pillow-accurate resize. This is the third tag on llama.cpp's semver track (after v0.1.2 and v0.2.0), which now runs in parallel with the usual bNNNN build tags.
Source
↳ Follow the thread