NaceAI open-sources Drex decision models, hitting SOTA 67.52 on Decision Index
nischay_twt · x · 2026-11-06
NaceAI released a 50-page technical report and fully open weights for Drex, a family of four decision models (8B, 9B, 35B-A3B, 27B) that return calibrated probability distributions over user-defined options in a single forward pass, no autoregressive decoding.
- Training: SFT plus their in-house RLAF recipe, blending verified outcomes with a reasoning judge under a KL constraint; ablations show RLAF lifts accuracy and consistency losses cut option-order flips from 6.3% to 4.2%. v1.5 trained on 1.15M verified decision records across 11 domains.
- Results: Drex Max sets SOTA among decision models on Decision Index 0.4.1 (67.52 vs. 60.70 for Jev 1.14.0), holds 66.60 on private tests, ranks second on Artificial Analysis's independent index, and is among the most-used decision models on OpenRouter.
- Weights and inference code: github.com/nace-ai/drex-decision-models
More from Models
- Google Ultra subscriber blocked from Antigravity by region detection — lxfater · 2026-11-06
- Palisade's Jeff Ladish: skip the 40 benchmarks, ask if agents finish real work — FinanceYF5 · 2026-11-06
- AA Index 31: 594B total / 17B active params, Apache 2.0 open source — AIFlow_ML · 2026-11-06
- Moonshot's Kimi K 3.1 Pro surfaces: 2.8T-param MoE, weights on Hugging Face — thursdai_pod · 2026-11-06
- Tencent Hunyuan Hy4 hits limits on day one, WorkBuddy rushes 10x compute expansion — TencentHunyuan · 2026-11-06
- Users say Claude Opus 4.8 is 'so stubborn even its own classifiers can't stop it' — repligate · 2026-11-06