l-forster/open-jet v0.5.4 was released on October 30, 2026. It is an offline coding agent for local models, a project that sets up a model runnable on the GPU with a single command. This version added a low-bit KV cache with attention handling, server interfaces, and installation and benchmark documentation. The release notes also state that graphics memory usage was reduced, and the source of the changes is the GitHub v0.5.4 release.
This project is intended to set up, in one place, an environment for doing coding work with local models. The repository description says that Qwen3.8-27B runs up to 1.9 times faster than llama.cpp on the RTX 30·40·50 series with 24GB or more of memory. That figure is a claim limited to the GPU, model, and comparison target stated in the description. A low-bit KV cache is a way of storing information about previous tokens with fewer bits during inference, and this release also covered the related handling and the server interfaces.
The release mentions a reduction in graphics memory usage but does not state the size of the reduction. The repository description's maximum 1.9-times performance claim also concerns a specific GPU, model, and comparison target, so there is no basis for reading it as performance in other environments. This update can therefore be seen as a change that fills in the memory handling and usage path of local inference, but the actual magnitude of the performance improvement is hard to judge from the information given.