Introducing Drex, a small model that decides instead of writing

Most of what enterprise software asks a model to do is not writing. It is deciding. Which queue does this ticket belong in. Is this clause a problem. Should this payment be held. Which of these four tools should run next. We have spent three years answering those questions with generative models, which means asking something that produces prose to produce a label, then parsing the label back out and hoping it is one of the options we offered.

Today we are releasing Drex by Nace.AI. It does not write. You give it a state and a set of typed questions, and one forward pass returns a probability for every option you supplied. There is no token stream, no chain of thought, and no JSON to repair. It cannot invent an option you did not give it, because it never generates anything at all.

It leads the Decision Index

The Decision Index is a public leaderboard of forty benchmarks that score decisions rather than prose, each corrected for chance so that guessing at random scores zero. On its 0.2 edition Drex scores 51.73, ahead of Jev 1.13.0 at 51.67 and of the strongest community entry, AutoJev-27B, at 50.94.

It is also the smallest model in the top ten. Drex has under 6B parameters; every other top-ten model with a published size has between 12B and 36B, and the strongest other model under 6B on the board scores 36.71, fifteen points lower. (Sizes are total parameters. Mixture-of-experts entries use only part of their weights per request, but still have to hold all of them in memory.)

It wins twenty-two of the thirty-nine benchmarks scored for it so far. The largest margins are on chord recognition, causal inference, sarcasm, multi-hop fact checking and contract reasoning. It trails on graduate-level science, broad subject knowledge and multi-step reasoning puzzles, which is the trade a small decision model makes and, we would argue, the right one: you are not asking your claims pipeline about quantum chromodynamics.

That shape is the actual finding. Drex is strong where it has been trained on the kind of decision in front of it. Your decisions are the ones that matter to you, and they are the ones nobody else has trained on.

It beats Jev at board games

We also played Drex against Jev 1.13.0 at eight board games, using the Kaggle Game Arena harness and OpenSpiel rules. Every legal move is offered as an option and each move is a single pass over the board, with no look-ahead. Each game has sixteen openings, each played from both sides, for 256 games in all. Drex won 117 to Jev's 92, with 47 draws, and came out ahead in five of the eight: checkers, Four in a Row, chess, Clobber and Nine Men's Morris. Jev took Reversi, Dots and Boxes and Lines of Action. The per-game results are on the product page.

Built to sit inside the loop

A generative model pays for every token it writes. Drex writes nothing, so a decision costs one pass over the state no matter how many options it scores. A single H100 sustains around 25 decisions per second.

To see what that makes possible, we pointed it at things that are not enterprise software at all. Drex plays a Doom deathmatch, choosing a target, whether to fire, where to move and which weapon to hold on every game tick: 727 consecutive decisions with no errors and no search tree. It commands a Terran base in StarCraft II and a fleet in a space battle, one call per loop. And a thousand simulated residents of a small town each decide what to do next, every simulated hour, for three days.

The recordings are on the product page, with Drex's and Jev's answers to the same cases side by side. To send it your own case, create a free API key.

Where it runs

Drex fits on a single accelerator, so it runs in your VPC, at the edge, or on-premises, as well as on Nacecloud. And because the gains come from training on decision data, it can be tuned on your own labelled outcomes, with the weights shipped to you. That is the same ownership model as the rest of the Nace stack: you should not have to rent your judgment.

Read the full benchmarks and try it, or talk to us about tuning it on your decisions.