Nace.AI releases Drex, a non-generative decision model

Nace.AI has announced general availability of Drex, a small model that makes decisions instead of writing text. Users supply a state and a set of typed questions, and a single forward pass returns a probability for each option. The model has no token stream, no chain of thought and no JSON output. Because it generates nothing, it cannot return an option the user did not supply.

At launch, Drex scored 51.73 on edition 0.2 of the Decision Index, a public leaderboard of 40 chance-corrected decision benchmarks. Jev 1.13.0 scored 51.67 and AutoJev-27B scored 50.94. Nace acknowledges that the 0.06-point lead sits inside the 0.25 gap the board treats as a tie. A September 28 update reports that Drex 1.5 scores 58.28 on edition 0.2.1 and ranks first of 71 entries under 10B parameters, 0.37 points ahead of Jev.

Nace gives the model's size as under 6B total parameters, with about 3B active per token. It says every other top-ten model with a published size has between 12B and 36B. Drex wins 22 of the 39 benchmarks scored for it, with its biggest margins in areas such as causal inference and contract reasoning. It trails on graduate-level science, broad subject knowledge and multi-step reasoning puzzles.

Across 256 board games against Jev, played with no look-ahead, Drex won 117, lost 92 and drew 47. Nace says one H100 sustains about 25 decisions per second. The model fits on a single accelerator and can be fine-tuned on a customer's labelled outcomes, with the weights delivered to the customer. Nace has not disclosed pricing or published any independent evaluation.