LLM Reddit Apr 14, 2026 2 min read
LocalLLaMA paid attention to this post because it looked like real engineering cleanup instead of another inflated speed screenshot. On April 13, 2026, the author said a stock-MLX baseline for Qwen3.5-9B at 2048 tokens improved from 30.96 tok/s to 127.07 tok/s, with 89.36% acceptance and the full runtime released as open source.