1000+ tokens/sec on H100s, a 5–10x speedup over typical autoregressive models — and without needing exotic hardware like Groq or Cerebras - impressive