Nace AI's open Drex models return answer probabilities without generating text
Nace AI's open Drex decision models return option probabilities without generating text, and a paper reports the 27B Max model scores 67.52 on Decision Index 0.4.1.
Drex is a family of four open decision models from Nace AI. Instead of generating an answer, each model returns a probability for every option a user supplies. According to the paper, its largest model, 27B Max, scores 67.52 on the public Decision Index 0.4.1 suite, compared with 60.70 for Jev 1.14.0. The models take a state, which the paper describes as a document, record or agent context, optionally with images. Questions ask for a choice among options, a yes/no judgment, or a score on an ordered scale. Because the model only has to produce a distribution over fixed options, it does not decode text token by token. The authors argue this removes the output parsing, inconsistency, cost and latency they associate with long chain-of-thought reasoning in production.
The four models are 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), and 27B (Max). Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, which the paper abbreviates as RLAF, and Max also receives vision tuning. In the RLAF setup, a teacher's option probabilities are adjusted using the verified answer and then serve as training targets. The paper's ablations credit RLAF with 2.0 Decision Index points. A separate consistency loss addresses another weakness. When the order of options is shuffled, a model can flip its answer. According to the paper, that loss cuts the option-order flip rate from 6.3% to 4.2%.
The benchmark comparisons need careful reading. On Decision Bench v4, the paper says Max is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points. The new state-of-the-art claim applies only among the decision models the paper compares, a qualifier the abstract makes explicit. The DAIR.AI Academy summary also notes that Max is slightly overconfident. Its mean confidence sits 5 to 9 points above accuracy in four of five domains.
The paper says Max answers a typical request in 183 ms. By comparison, the figures are about 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5. These latency figures cover a single typical request, and the abstract does not describe the hardware or batching used to measure them. The comparison is best read as a per-request figure, not a general throughput claim.
The paper's materials are public. Its abstract says all four models are open source and links to a GitHub repository for the models and code. The arXiv listing shows a 51-page paper with 13 figures and 25 tables, dated 5 Nov 2026.
What changed
Nace AI releases four open Drex decision models that return option probabilities for user-supplied options instead of generating answer text.
Type: New capability · Applies to: 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), 27B (Max)
Why it matters
Teams running classification or routing decisions can get calibrated option probabilities in one forward pass, which may cut latency and cost compared with generating answers.
What you can do
Benchmark an open decision model on your own fixed-option workflows against your current model for latency and answer stability.
How the story developed
Adds details from the Drex paper on arXiv, via DAIR.AI: the four model sizes, Decision Index scores, calibration and per-request latency.
Original source
Where we saw it
DAIR.AI (Elvis Saravia), Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works