I ran Qwen3.5-0.8B in pure Assembly

On an AMD EPYC 7B12, it generated 6.51 tokens/sec on one CPU thread. The Transformers reference reached 4.34. All 32 generated tokens matched in both runs. Code: github.com/suayptalha/qwen3_5-assembly X Post: x.com/suayptalha/status/2119254536283906205 Tokenization, model inference and sampling all run in Assembly. No libc or Python at runtime; Python prepares the tokenizer tables at build time. Benchmark: 19 input tokens, greedy decoding, 2 runs. Assembly reads BF16 weights; Transformers loaded them as FP32 and used PyTorch fallback kernels.