Diffusion LLM goes production: Augment Code's Mercury 2.5 switch cuts latency 82%, cost 90%

cen6wkf · reddit · 2026-09-25

Augment Code quietly swapped its production coding-agent backend in September to Mercury 2.5, Stefano Ermon's diffusion LLM — a smaller model on a non-autoregressive architecture — cutting latency 82% and cost 90% in a shipped product. Artificial Analysis independently measured 770 tokens/s vs Inception's claimed 1,107. The post argues autoregressive inference is sequential and memory-bound while diffusion generates token blocks in parallel, a real fit for GPU bottlenecks. It also spots an open business gap: no neutral benchmark runs diffusion vs autoregressive stacks side-by-side on your own traffic, and proposes a concrete plan using open-weight DiffusionGemma 26B-A4B vs Gemma 4 26B-A4B on one rented H100 ($3.41/GPU-hour) via vLLM, noting governance frameworks regulate model outputs but not the engineers whose expertise just got bypassed.

Original post →

More from coding & agent

coding & agent channel →