Benchmark: decision models run multiple times faster than LLM equivalents, single forward pass

mervenoyann · x · 2026-11-18

The author benchmarked LiquidAI's LFM3 against d1 (excluding warmup, on Llama), showing decision models are several times faster than their LLM equivalents for multiple options: an LLM needs prefill + decode for JSON structured output, while a decision model does a single forward pass. He also notes a quirk in the implementation — it prefills multiple times per question (unlike clef) — so the speedup will be even larger for the latter.

Related event: Liquid AI's Decision Model Beats LLMs at JSON Generation by 2.7-3.6x(2 posts)→

Original post →

More from Models

Models channel →