kekzl/imp released v0.60.0 on October 31, 2026. The project is a tool for running large AI chat models locally on NVIDIA RTX 50 or RTX PRO graphics cards. This version applies its own math kernel to long-prompt processing for some models. Release testing reported a 6.0% improvement in Gemma-4 long-prompt processing and a 5.4% improvement in Qwen3-Coder short-prompt processing. Changes to embedding and NVFP4 memory placement and to the API are also included, and it states that there are no compatibility changes in the default configuration.
According to the project description, a single RTX 5090 generates about 70 words per second with Qwen3.8-27B, which is 1.4 times faster than vLLM. It also says that models nearly four times larger than the graphics card's memory can be run. It can be used together with apps made for ChatGPT or Claude, and it starts with a single Docker command. So this release is an update that works on prompt processing, memory placement and the API together in a local model execution environment.
However, the own math kernel is applied only to some models, and the reported improvement figures are release test results for specific prompt types. The provided release summary does not state the measurement conditions or the comparison baseline. The details of the API changes also do not appear in the summary. Along with the statement that there are no compatibility changes in the default configuration, the effect of the improvements is better judged within the range of the applicable models and prompts.