Developers introduced the compact MoE model Edge0-35B-A3B-preview

The team of developers released Edge0-35B-A3B-preview — a new language model of the Mixture-of-Experts architecture with a total of 35 billion parameters. The project gained over 1200 stars on the Hugging Face platform just 44 hours after publication.

The key feature of the model is optimization for operation on devices with limited resources. The system consumes less than 3 GB of active RAM, allowing it to run on consumer equipment. When using the edge0 framework, the model demonstrates a generation speed of around 15 tokens per second.

During testing, the developers recorded a loss of response quality at the level of 3.9 points compared to full-size analogs. Such indicators are achieved due to a specific configuration of experts that minimizes computational load while preserving the model's logical abilities.

The release is focused on local usage scenarios where low VRAM consumption and high response speed are critical. The model is available for download in the project repository, where instructions for integration into existing inference pipelines are also provided.