Liquid d1-2B decision model answers 19-30x more questions/min than same-size LLMs
mervenoyann · x · 2026-11-20
zekiworks benchmarked Liquid AI's decision model on the same GPU (RTX PRO 6000), with mervenoyann amplifying and explaining why it wins:
- Liquid d1-2B answers an estimated 5,136 English reading questions per minute — about 19x Liquid LFM3-1.7B and 30x MiniCPM5-2B/Qwen3.5-2B — while tying them on accuracy (65.6% vs 67.8%/67.3%) across 547 questions.
- With thinking off, Liquid LFM3-1.7B is the most accurate chat model at 71.8%.
- The reason: LLMs need prefill + decode to emit JSON structured output, while decision models do a single forward pass, yielding order-of-magnitude throughput gains for structured-output tasks.
More from Infra
- Cactus Compute publishes 12 papers in 2026, 5 at NeurIPS, focused on on-device efficiency — joemeno · 2026-11-20
- Modular rebuilds MAX's Module API with PyTorch-like syntax and eager mode — clattner_llvm · 2026-11-20
- e-means lands across tusKANNy stack, cutting vector index clustering training time by 1.5-5x — CShorten30 · 2026-11-20
- AWS Lambda doubles function timeout to 30 minutes for on-demand Event Functions — DavidWells · 2026-11-20
- Microsoft taps Intel EMIB packaging for Azure silicon, locks in Diamond Rapids supply — Beth_Kindig · 2026-11-20
- Polymarket puts DOJ criminally charging Nvidia this year at just 5% — Polymarket · 2026-11-20