⬢github C++ · 898 ★ +90 since we first saw it · pushed 21 h ago · MIT
mw00/project-maya
Project Maya - run GLM-5.3-Flash (321B MoE) privately on your own GPU(s), from one card to 16, NVIDIA or AMD, Linux or Windows. Engine, server, dashboard, OpenAI- and Anthropic-compatible API, quant tools.
Project Maya is a C++ engine and server for running the 321B-parameter GLM-5.3-Flash MoE model privately on local hardware, from one GPU up to 16, NVIDIA or AMD. It keeps frequently-used experts in VRAM with the rest in RAM or SSD, auto-tunes to your machine, offers quantized model variants, a web chat dashboard, and OpenAI/Anthropic-compatible APIs with batched requests.
Why now: It's trending as a new, heavily-starred repo (808 stars in about a week) that makes a huge 321B model runnable on consumer hardware with benchmarked speed and accuracy claims.
Who it is for: Hobbyists and developers with gaming or workstation GPUs who want to run a frontier-size LLM locally without sending data to the cloud.
cudaglmglm-5llmllm-inferencelocal-llmmixture-of-expertsmoeopenai-apiquantization
Stars over our 115 snapshots: 808 to 898, since 2 d ago.
Where people talked about it
- ⬢github new repos, most starred 10 min ago
- ⬢github new repos, most starred 13 h ago
- ⬢github new repos, most starred 1 d ago
API: https://socialmediatrends-api.osmike.com/v1/repos/mw00/project-maya