zephel01/abyss is an inference engine aimed at running DeepSeek V4 Flash and V4.1 Flash family mixture-of-experts models on a single consumer graphics card. The latest recorded release tag is v1.3.54. What changed in this version, however, is not given in the provided information. The README explains that the project was forked from Project Maya.
A mixture-of-experts model (MoE) is a scheme that picks and uses only the parts it needs from among several experts. Abyss describes a tiered execution model that keeps frequently used experts in graphics card memory and loads the rest from fixed memory and NVMe storage. It targets Linux, supports CUDA, and its ROCm support is experimental. It also provides a browser chat screen and OpenAI- and Anthropic-compatible APIs. It is written in C++, licensed under MIT, and has 18 stars. The README holds benchmark tables for several graphics card and model configurations, an explanation of how it works, and a manual.
The benchmark tables present hardware and model configurations together, so the conditions have to be checked when reading the numbers. ROCm support in particular is marked as experimental. What this repository shows, then, is a tiered structure and support path for single-graphics-card execution, and the measurements are best read as limited to each set of conditions.