Breaking01 / AI
Drex Max decision model scores 67.52 on Decision Index 0.4.1
The score and response times of the open model, which returns only a probability for each option, are the authors' own figures and have no independent replication.

Nace.AI reported on 5 Nov 2026, in the not yet peer-reviewed paper Towards Foundation Decision Models: Drex, that its Drex Max model scored 67.52 on Decision Index 0.4.1 (public suite); no independent replication exists.
The paper names inconsistent answers, cost and latency as the problems long step-by-step reasoning creates in production services. Its alternative is the decision model, which, the researchers reported, takes a question and a set of options and returns a probability distribution over those options without generating text.
The Drex family has four models, according to the paper: DLM (8B parameters), v1.5 (9B), v2.0 (35B, with 3B active) and Max (27B). The input is a state, such as a document, a record or agent context, together with questions about it, and answers come as a choice among options, a yes/no judgment or a score on an ordered scale. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback (RLAF), and Max also has vision tuning.
On Decision Bench v4, the paper reported that Max is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points (same evaluation). Response times are 183 milliseconds for Max, about 2.0 seconds for GPT-6 Sol and about 3.5 seconds for Claude Opus 5.5 (typical request).
Availability. The GitHub repository lists the Max, v2.0 and v1.5 weights under the Nace.AI Open RAIL-M licence and the DLM weights under CC BY-NC 4.0. Max's affine Q4 checkpoint runs on Apple silicon Macs with 24 GB of memory or more.
The paper reported that the Max score is ahead of the 60.70 recorded by Jev 1.14.0 (Decision Index 0.4.1, public suite). The authors limit that top-score claim to the decision models they compared.
Neither the scores nor the response times have been independently replicated, and the abstract does not disclose the hardware used to measure the 183-millisecond response time (Max, typical request).