What is q36?

q36 is a specialized inference engine built to run one specific large language model, Qwen3.6-35B, as fast as possible on one specific graphics card, the NVIDIA RTX 5090. Instead of trying to support many different models and GPUs the way general purpose tools like llama.cpp or vLLM do, q36 is written entirely in C and CUDA and optimized narrowly for this one pairing, on the idea that a purpose built engine can outperform general tools on the cases they were not specifically tuned for. According to benchmarks in the project, q36 processes prompts faster than llama.cpp at every context length tested, and generates new text faster in nearly every case as well, with one exception at very long context depths. It achieves this partly through custom CUDA code for both the attention layers and the recurrent memory layers the model uses, and through techniques like shrinking the memory used to store conversation context and instantly saving and restoring conversation state from system memory in a few milliseconds. It ships with a command line chat interface, a server that speaks the same protocol as OpenAI's API so existing tools can connect to it, a benchmarking tool, and a way to measure how accurate the model's outputs are compared to llama.cpp on a standard test set. Using q36 requires specific hardware, an RTX 5090 or a similar Blackwell generation NVIDIA GPU, plus a specific CUDA toolkit version installed on Linux, and downloading the Qwen3.6 model weights separately from Hugging Face. This is not a general purpose tool for casual users. It is aimed at people already running large language models locally who have the exact hardware it targets and want the fastest possible performance from it. The project has 31 stars on GitHub, and its license is not stated in the material reviewed.