mw00/project-maya
Project Maya says it can run GLM-5.3-Flash, a 321B-parameter model, locally by tiering experts across GPU, RAM, and NVMe storage.
The repo describes Maya as a purpose-built engine for GLM-5.3-Flash, with about 18B active parameters per token and context up to 1M tokens. It lists quantized model options from roughly 95GB to 156GB, claiming 97.7% to 99.2% of the FP8 model’s zero-shot accuracy. Reported speeds range from 118 tokens/s on four RTX 4090s to 30+ tokens/s on a single RTX 5090, with broader user benchmarks included. It ships a browser chat, live monitor, image input, and OpenAI- and Anthropic-compatible local APIs. Source: GitHub · LLM repos' note.
The repo describes Maya as a purpose-built engine for GLM-5.3-Flash, with about 18B active parameters per token and context up to 1M tokens. It lists quantized model options from roughly 95GB to 156GB, claiming 97.7% to 99.2% of the FP8 model’s zero-shot accuracy. Reported speeds range from 118 tokens/s on four RTX 4090s to 30+ tokens/s on a single RTX 5090, with broader user benchmarks included. It ships a browser chat, live monitor, image input, and OpenAI- and Anthropic-compatible local APIs. Source: GitHub · LLM repos' note.
score 5