kekzl/imp is a C++23 LLM server that targets a single NVIDIA Blackwell GPU. The latest release is v0.59.2. The provided information does not separately list what changed in this version. It runs in a single container and supports OpenAI- and Anthropic-compatible APIs along with tool calling.
The server focuses on native 4-bit arithmetic on Blackwell (sm_120a), tensor cores, and low-precision compute paths. Its input formats are GGUF and NVFP4. The README publishes measurements taken on an RTX 5090 along with the comparison baseline. Running Qwen3.8-27B as a single user yields 96.3 tokens per second. In the same measurement, it is recorded as 1.38x faster than vLLM without speculative decoding and 1.23x faster than llama.cpp using the same GGUF. It also lists a figure of serving Qwen3.8-Flash-Next, 123.6 GiB in size, at 92.1 tokens per second on a 32 GB card.
The applicable scope is limited to single-Blackwell-GPU environments. The README explicitly states that environments other than a single GPU are not supported, and the targets are the GeForce RTX 50 and RTX PRO series. Therefore, the published figures and features do not apply as grounds for environments using other GPUs or multiple GPUs. The comparison results should likewise be read within the conditions stated in the README, and the supported range of this server is on the narrow side.