blog

Engineering notes from building hipfire — RDNA kernel work, speculative decode, falsified hypotheses, perf checkpoints.

• 2026-10-07

  hipfire 0.4.1: faster local agents on Strix Halo

  hipfire 0.4.1 runs Qwen3.8-Flash agents locally on Strix Halo: sampled MTP with penalties, a multi-session prompt cache and HF-exact tokenization.