helldez/bigmoeonedge released 0.40.1 on 2026-11-06. It is a project that aims to run MoE (Mixture of Experts, a structure that computes only some experts per token in order to reduce cost) models larger than RAM on top of llama.cpp, and the repository description states that it runs frontier-class MoE losslessly even on phones with around 12 GB. This commit raises the bundled llama.cpp submodule to the upstream tag v0.9.0 (2026-11-04, cd959ba), catching up 164 commits from the 2026-10-28 baseline 8ca4060. The fork branch bmoe/expert-ready-hook-0411 merged only the single expert-ready-hook commit into that tag without conflicts.
According to the CHANGELOG, what v0.9.0 brings is new architecture support and tensor-type changes. The newly supported architectures include Clover-35B-A3B, Mistral Large 4, Metis-1.6, DeepSeek V4.1 Flash, the LFM3 family, Nemotron Bolt 74B-A9B, and Kolibri-1, and the FP8 E4M3 tensor type and prefill-related work for Gemma 4 E2B/E4B came in together. No engine-side changes were needed; the expert-ready-hook still fires ahead of upstream's tiled mul_mat_id path, only the FP8 E4M3 type was added to the CPU seam file, and the scheduler's eval callback area was left untouched. The patch size is 30 files with 21 lines added and 9 lines deleted, and the versions in the engine, the app, and the CHANGELOG were aligned to 0.40.1.
Validation is limited to what the commit message discloses. All 41 host tests, including the overlap gate, passed, and there is a record across two runs that on an actual 35B model --auto produced the same greedy text as the previous pinned state with the same CPU time per token. The list of new architectures is likewise only what the tag provides, not results measured on this engine. No throughput or speed figures are presented, and this is a project-side report rather than an independent measurement. Whether the new architectures actually deliver quality and performance is still too early to judge.