blockwayz/agens-engine is a C++ inference engine that runs the Agens Volundr model on a single consumer graphics card and serves it through an OpenAI-compatible chat API. The repository was created on October 17, 2026 and has no public releases. The only things that can be checked right now are the repository description and the features and run profiles listed in the README.
What this engine targets is planning the weights and a long context cache together inside a GPU memory budget. It uses layer-wise mixed precision that stores the weights of each linear layer at 2-bit, 3-bit and 4-bit, and offloads the KV cache holding the keys and values of generated tokens to host memory. It also supports speculative decoding that uses an auxiliary draft model and an MTP head (drawing draft tokens first and validating them in one pass), execution that splits the model across multiple graphics cards, and image input. The API is OpenAI-compatible. The 16GB profile requests a context of up to 262K tokens, but the actual capacity varies with GPU memory and the KV pool, and it is designed to abort execution when the memory budget is exceeded. The README lists run profiles and limitations for each graphics card memory size.
That said, the server, the CUDA kernels and the parity tests are confirmed, but the public performance claims require separate verification with reproduction material. The license is marked NOASSERTION, so the usage terms are unclear. There are 7 stars; on a 10-point scale, novelty is 8.0, technical is 8.0, quality is 7.0 and momentum is 1.8. The design is concrete, but the performance evidence is thin. Because the momentum score is low and the repository has only just been created, it is close to a project worth checking right now for people who want to run Agens Volundr directly on a consumer GPU.