nipegun/moe-gguf-server added the repository wholesale in a commit on November 3, 2026 (6a9e2d0). 671,979 lines came in and 0 lines were deleted, and the commit message is “Tokens... My precious!”. This project is a server that runs MoE (Mixture of Experts, a structure in which multiple sub-models are handled within a single model) models in GGUF format locally, and provides a web chat and an HTTP inference API.
The README mentions llama.cpp and FreeToken and states that it aims for the fastest possible inference speed in MoE models. The backend is CMake-based C++ code, and it supports Vulkan (AMD·Intel·NVIDIA), CUDA (NVIDIA), and ROCm (AMD) build paths respectively. Toolchain files for arm64 Linux, macOS, and Windows, as well as for riscv64 (SpacemiT Linux), were included as well. The GGUF architectures read by mgs are enumerated as qwen35moe (Qwen3.6), glm-dsa (GLM-5.3), kimi-k3 (Kimi-K3), deepseek41 (DeepSeek-V4.1-Flash), minimax-m3 (MiniMax-M3), flint (Flint). backend/common contains chat parser and change-point analysis modules such as chat-peg-parser and chat-diff-analyzer, and the README also provides 15 language versions.
The LICENSE added in the same commit is the PolyForm Noncommercial License 1.0.0. Personal use and use by noncommercial organizations (charities, educational institutions, public research/safety/health agencies, government agencies) are permitted, but commercial purposes are not covered, and sublicensing and assignment of the license are also prohibited. The benchmark figures backing the “fastest inference” goal are not in this commit, and the README also states that screenshots have not been added yet. Distribution is guided by a Debian 13-only installation script that requires root privileges, and ROCm builds require HIP 6.1 or higher, so they cannot be built as they are with only Debian 13's default packages (HIP 5.7). In short, it is a situation where performance evidence cannot be verified, with the conditions that block commercial use layered on top.