shawnshekari/strixite is a HIP-based LLM inference engine built to run Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151). The latest release is v0.3.1, and the provided information does not list what changed in this version. The repository description states that it runs at roughly 50 tokens per second with up to 492k context. The README explains that it was written from scratch with HIP instead of relying on llama.cpp or ggml, with the kernels implemented directly. This is an approach tailored to AMD Strix Halo and a specific model combination rather than one broadly supporting general-purpose engines.
The engine targets local inference that keeps a long context while generating model responses. According to the README, it supports 512k context on a single Strix Halo with 128GB of memory, with prompt processing at 1,370 tokens per second. It also lists a RAM/disk prompt cache and multi-token prediction that checks up to 7 tokens at a time. It is written in C++ and licensed under AGPL-3.0, and the repository has 56 stars. The README also mentions an MoE architecture with 512 experts and a 51B-parameter n-gram table read from SSD.
That said, the performance figures are the repository's own claims, and there are no public measurement scripts or benchmark result files. The README's 19/19 record on the terminal-bench-mini core suite alone makes it hard to reproduce the speed at 492k context or generalize it to other environments. So the specific changes in v0.3.1 and any independent verification of performance cannot be confirmed from the provided information, and it is more reasonable to treat this as an experimental engine for a particular hardware and model combination.