mw00/project-maya released v1.1.41. It is a project built to run GLM-5.3-Flash (321B MoE), DeepSeek-V4.1-Flash-1017 (748B MoE) and MiniMax-M3.1-Flash (428B MoE) on the user's own GPU. This release includes 36 community contributions. According to the release notes, in multi-GPU environments the speed of reading the first long prompt right after startup is about 1.3x faster.
Beyond the model execution engine, the project also provides a server, a dashboard, OpenAI- and Anthropic-compatible APIs and quantization tools. It says it supports one to sixteen NVIDIA or AMD GPUs along with Linux and Windows. So the components needed to run models can be reviewed in a single repository, and setups that use the compatible APIs are an option as well. The release notes add that speed improved even on PCs with insufficient memory.
v1.1.41 also added an agent loop-prevention mechanism that keeps working even after turns change. However, the release description does not state the measurement environment or comparison baseline for the roughly 1.3x improvement. This figure can be seen as a result reported by the release, but no evidence is given to conclude that the same improvement occurs in other environments.