Nace AI open-sources Drex decision models: 27B rivals frontier LLMs at 183ms
omarsar0 · x · 2026-11-06
Nace AI released Drex, a family of four fully open decision models (8B, 9B, 35B with 3B active, and 27B Max), with weights and inference code public.
- How it works: given a state (document, record, or agent context) plus questions with fixed options, the model returns a probability distribution over options without autoregressive decoding—avoiding inconsistent outputs, parsing overhead, and latency.
- Results: Drex Max scores 67.52 on Decision Index 0.4.1 (vs 60.70 for Jev 1.14.0), is statistically tied with GPT-6 Sol and d3-turbo on Decision Bench v4, trails Claude Opus 5.5 by only 1.6 points, and ranks second among decision models on the independently run AI Digest index.
- Speed: 183ms per typical request vs 2s for GPT-6 Sol and 3.5s for Claude Opus 5.5, with stable outputs under option reordering.
- Training: supervised fine-tuning plus RLAF (RL from adjusted feedback); ablations show RLAF improves accuracy and consistency; vision tuning applied only to Max.
Related event: Nace AI Open-Sources Drex Decision Model Family(4 posts)→
More from Models
- Jeff v1.4: a 0.8B decision model with 34 open adapters, 90%+ accuracy, 2ms inference — Usual_Maximum7673 · 2026-11-06
- Qwen4 rumored to drop today, teaser image leaks — HankYeomans · 2026-11-06
- Claude fails 4 times to find a site owner's name on-page, then tells user not to subscribe — WisdomLoading · 2026-11-06
- Open-source decision model family Drex debuts, Drex Max tops decision benchmarks — arankomatsuzaki · 2026-11-06
- Fax-degraded table fools 7 of 16 top open models into misreading a child's dose — MaziyarPanahi · 2026-11-06
- BAAI's AREX-2: a 27B open-weights model that runs its own try-check-retry agent loop — Sam Witteveen · 2026-11-06