I gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes

Qwen 3.8 27B runs on a Lenovo ThinkStation PGX with Nvidia Grace Blackwell, achieving about 50 tokens per second using SGLang, NVFP4 and DFlash2 speculative decoding. In a reverse-engineering test, the model statically analyzed an ARM64 license-check binary, identified the hidden public verification key, and produced a working bypass after documenting the process.