Drex: open-source decision models that skip text generation, 183ms vs 3.5s

TheTuringPost · x · 2026-11-08

NaceAI released Drex, a family of four open-source decision models (8B DLM, 9B v1.5, 35B-with-3B-active v2.0, 27B Max) that skip autoregressive text generation entirely — given a question and options, they output a probability distribution over options, avoiding the inconsistency, parsing overhead, and latency of long chain-of-thought in production.

Architectures: three smaller models use a pointer architecture (pretrained LM + small head comparing option representations with the question; context encoded once and reused across questions). Drex Max is a dense multimodal model that reprocesses context per question with bidirectional attention over up to 255 options jointly.

Training: SFT on labeled decisions across 11 domains, then RLAF (RL from adjusted feedback mixing verified answers with a teacher model's probability estimates) for confidence calibration; Max adds a vision stage.

Results: 183ms per query vs 2s (GPT-6 Sol) and 3.5s (Claude Opus 5.5); new SOTA on Decision Index 0.4.1 public suite (67.52 vs 60.70 for Jev 1.14.0); statistically tied with d3-turbo and GPT-6 Sol on Decision Bench v4, 1.6 points behind Claude Opus 5.5; consistency training cut answer flips from 6.3% to 4.2% under option reordering.

Paper and code are open.

Related event: Nace AI Open-Sources Drex Decision Models That Skip Text Generation(11 posts)→

Original post →

More from Models

Models channel →