Tools
llama.cpp 0.4.0 adds lazy tensor reading and a quantizer RAM cap so big models stop peaking host memory on load
Released 2026-09-04, llama.cpp 0.4.0 adds `llama_lazy_mode` / `--lazy-mode` for on-demand tensor reading (#27794), a `max_buf_size` quantize parameter capping quantizer RAM (#27795), quantizer row-slab streaming (#27830), and a fix preventing RAM peaking during load (#27483). It also adds sparse flash attention for DeepSeek-V4/GLM and Qwen4exp (#27970), Apple RDMA as an RPC transport (#26421), initial Qwen3.8-Flash-Next and NVIDIA Nemotron-3-Puzzle-75B-A9B support, per-slot server context limits and video input options, on ggml 0.23.0.
Source
↳ Follow the thread