Nace makes Drex decision model generally available
Nace says Drex is now generally available and ranks ahead of Jev on the Decision Index 0.3.1, which supports its case that small non-generative models can handle routine production choices.
Nace has made Drex generally available. The company says in a new explainer that the model ranks ahead of Jev on the public Decision Index 0.3.1 and uses about 3B active parameters per token.
As we reported earlier, Drex does not generate text. It reads a case, a question and a fixed list of options, then returns a probability for each option in one forward pass. Nace says this means it cannot produce answers outside the given options, it bills only input tokens, and several questions about the same case can share one pass.
The post also gives figures from Nace's technical report for the vision-capable Drex Max:
- Latency: Drex Max answers a typical request in 183 ms. Nace puts GPT-6 Sol at about 2 seconds and Claude Opus 5.5 at about 3.5 seconds.
- Accuracy: On Decision Bench v4, Drex Max is statistically tied with GPT-6 Sol and 1.6 points behind Opus.
- Calibration: When Drex Max is at least 93% confident, it is right 97% of the time.
- Consistency: Shuffling the order of the options changes 4.3% of its answers. Nace says most of those were answers the model had already marked as uncertain.
These are Nace's own numbers. The post does not include the independent Decision Index scores behind the ranking against Jev, which Artificial Analysis maintains.
Nace says decision models do not replace LLMs for open-ended output, for questions without a fixed list of answers, or for answering from memorized knowledge. The Drex family has four open models, with weights on GitHub and Hugging Face. Drex Max is served through the Drex API.