jaedync/crucible-strix-halo is an inference engine that runs EXL3 checkpoints of Qwen3.8-Flash-Next on AMD Strix Halo systems. Its targets are the Ryzen AI Max+ 395 APU and Radeon 8060S graphics. There is no public release version. The README publishes a performance measurement table along with the measurement method and the reproduction conditions.
Rather than wrapping another inference engine, this repository uses its own ROCm/HIP kernels written in C++20. It provides chat, streaming responses and function calling through a local OpenAI-compatible server, and it is described as supporting accurate speculative decoding. Speculative decoding is a method that predicts the next token in advance and then verifies the result to generate it. The README includes a throughput table by context length and model compression ratio, as well as results from running multiple sessions at the same time. The license is Apache-2.0.
However, the target of this engine is limited to specific GPU architectures, including Ryzen AI Max+ 395 and Radeon 8060S. The author also defines it as research software rather than a general-purpose backend or an operational service. The throughput figures are measurement results published by the author, so they do not represent performance in other environments or with other engines. Therefore, it is more accurate to view it as an implementation and measurement material for that particular hardware rather than a general-purpose inference tool.