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, OpenA
What is 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.
Overview of project-maya
What it is
Project Maya is a software project that runs the GLM-5.3-Flash 321-billion-parameter Mixture-of-Experts (MoE) language model on private, user-owned GPU hardware. The project provides an engine, a server, a web dashboard, and an OpenAI- and Anthropic-compatible API. It supports one to sixteen GPUs, both NVIDIA and AMD (experimental), on Linux or Windows (experimental). The project is created and led by Peasant Smith (@mw00), co-maintained with @needmorevram. The license is MIT.
Key features
- Runs a 321B MoE model privately on user hardware without sending data to external servers.
- Dynamic expert offloading: keeps active experts on GPU, next in RAM, and the rest on an NVMe SSD.
- Self-tuning: measures GPU, CPU, RAM, and SSD performance and adjusts configuration automatically.
- Supports up to 8 simultaneous requests from multiple users or agents using the built-in parallel decoding feature.
- Includes quantized model variants (Maya-S, Maya-S24, Maya-M, Maya-M-Derisked, Maya-L) that preserve 97.7%–99.2% of the original FP8 model's zero-shot accuracy.
- Provides a browser-based chat interface and live monitor, support for images, thinking levels, tool calls, and a one-command setup script.
Use cases
- Running a large-scale AI model locally for privacy-sensitive applications, such as analyzing personal documents or data that cannot leave the user's machine.
- Powering coding agents or multi-session AI tools that require handling several requests concurrently.
- Experimenting with or fine-tuning a 321B-parameter model on consumer or prosumer GPUs without access to a datacenter.
Getting started
See the repository README for setup instructions. The project includes a one-command setup script and offers selection of quantized model variants (Maya-S, Maya-S24, Maya-M, Maya-M-Derisked, Maya-L) during installation. Users can run benchmarks with ./maya.sh --bench and report results with ./maya.sh --report.
License and pricing
The project is licensed under MIT. The pricing type is open source, meaning the software is free and publicly available.
Based on the project's public README and repository metadata.
