llama.cpp v0.3.0 — first tagged stable release of the b10621 nightly; dots3-note multimodal, GLM-4.5-Air MTP, DeepSeek 4 tensor-split, ggml v0.22.0
ggml-org
Tagged stable release on Aug 25, built from nightly b10621. New model support: dots3-note with a new DSA-ISWA KV cache type (#27060); MTP (multi-token prediction) for GLM-4.5-Air (#26534); DeepSeek 4 `-sm tensor` mode (#26490) and rollback fix with multiple sequences (#26756); bailingmoe3 DSpark (#27508); mamba2 in/out projections flattened to dispatch GEMM instead of GEMV (#27513); RoPE `set_offset` used in deepseek2/4, dflash, minicpm3, plm. mm: WebP via ffmpeg (#27520); videos with moov atom at file end fixed (already in b10608); Pillow-accurate resize (#27594); dots3-note vision and audio (#27524). Server: `LLAMA_SERVER_SLOTS_N_DIFF` debug knob (#27600); slot fitting moved to common `fit` taking `n_streams` into account (#27496); common `json.h` abstraction (#27511). UI: tabbed navigation for chat conversations (#27263). ggml v0.22.0: tensor-split support for multi-backend (meta backend); Metal per-op source split for parallel compile; `ggml_clamp` made non-in-place; new ops `POOL_1D`, `PAD_REFLECT_1D`; Q2_K SYCL kernels; MoE bias fusion on OpenCL; assorted CUDA/Metal/SYCL/Vulkan/OpenCL/WebGPU fixes. Notable commits since v0.2.0: b10625-b10631 pre-releases added ROCm Ubuntu 24.04 CI, Apple RDMA RPC transport, Metal OOM null-check, kleidiai reworked integration.
Why it matters
v0.3.0 is the official stable tag after the b10603-b10615 series already covered yesterday. Standalone novelty in this release is the dots3-note multimodal family and ggml v0.22.0 (Metal compiles a lot more in parallel, so build times and binary size change meaningfully).
Importance: 3/5
llama.cpp tagged stable release with new model family + ggml major bump