Project Maya v1.1.16 enables local execution of 748B DeepSeek model on consumer PCs

Japanese tech outlet Impress Watch reported on October 20, 2026, that the open-source AI inference engine Project Maya released version v1.1.16, officially adding support for running the massive 748-billion-parameter (748B) large language model DeepSeek-V4.1-Flash-1017 locally on personal computers.

Project Maya aims to break the dependency of ultra-large models on data-center-grade hardware. Its core technology combines custom quantization algorithms with a specialized engine: the engine dynamically allocates 'expert' modules within the model to different storage tiers based on usage frequency during conversation. The most frequently used parts reside in NVIDIA GeForce series GPUs, less frequent parts are stored in system memory, and the remainder is loaded onto NVMe SSDs. This tiered strategy allows the model to achieve high-speed, private local inference on consumer PCs while maintaining performance close to the original.

According to self-reported test data from the project team, when running DeepSeek-V4.1-Flash-1017 with the default quantized version 'Maya-S', the consistency rate of the next generated token with the original model's output is 92.6%. The update also supports multi-GPU environments and introduces a feature to acquire DSpark drafters, allowing users to further accelerate draft processing if their hardware conditions permit.