mw00/project-maya v1.1.45 was released on October 31, 2026. This version fixes the problem where long prompts failed with invalid memory access when DeepSeek-V4.1-Flash ran on a single GPU. According to the release notes, the fix cleans up memory regions that overlapped with experts being loaded in the background. The source of the change is the v1.1.45 release of mw00/project-maya.
mw00/project-maya is a project for running GLM-5.3-Flash (321B MoE), DeepSeek-V4.1-Flash-1017 (748B MoE) and MiniMax-M3.1-Flash (428B MoE) on your own GPUs. It targets GPU environments from one to sixteen cards, and the repository states that it supports NVIDIA and AMD GPUs along with Linux and Windows. Besides the engine, server and dashboard, the OpenAI- and Anthropic-compatible APIs and quantization tools are also within the provided scope. So it is not limited to running one specific model, but covers the components needed both to run models and to serve APIs.
The scope of the fix given in this release description is the error that occurred while processing long prompts for DeepSeek-V4.1-Flash on a single GPU. Nothing is stated about the same problem being resolved for other models or GPU configurations. It is therefore hard to read the v1.1.45 change as a fix covering every model and hardware combination.