Nace AI open-sources Drex, a diffusion-LM decision model

Nace AI has published the weights for Drex DLM, a decision model. Instead of writing a reply, it reads a document and returns a probability for each answer the developer defined in advance. The model card shows a support ticket from a customer who was charged twice, sent with three named questions: which team should handle it, whether the customer explicitly asks for a refund, and how urgent it is on a scale from routine to urgent. In the card's local run, the model gave billing a probability of about 0.96. A program can act on numbers like that directly, for instance by sending the ticket to billing above a set threshold, without having to interpret written text.

Drex accepts the same request format as Jev, the decision model from TypeSafe AI that Simon Willison described as a new shape of language model in earlier Intel coverage, and the card says software written for that format works unchanged. Nace built Drex on Efficient-DLM-8B, a diffusion language model from NVIDIA, and added a small scoring layer it calls a pointer head. The model reads the document in both directions, so each word is interpreted with the text around it. Each question then reads the document, and the pointer head compares the model's internal state at the question's decision marker with its state at the end of each option, turning those comparisons into probabilities. When a request fits the model's token budget, every question is scored in one pass through the model.

Nace's own results table puts Drex at 52.31 on version 0.2 of a Decision Index leaderboard, ahead of the next five entries as of October 2026. The lead is narrow and uneven. Drex is 0.38 points ahead of the second entry, a Gemma-4-31B-based model, and it posts the best scores of the six in knowledge and reasoning, in retrieval and classification and in arts and human taste. It has the lowest score of the six in tools and automation, 48.59 against 66.86 to 69.98 for the others, and also in language understanding. The card does not say how the index is built, so the table shows where Drex ranks on Nace's chosen comparison, not how it will handle a particular workload.

The announcement calls Drex open source, but the card licenses the weights under CC BY-NC 4.0, which bars commercial use, while the code is Apache-2.0. The card documents three ways to run it locally: Nace's Python server and its modified forks of llama.cpp and Ollama. The card says it was tested on Apple M5 Max and M5 Pro machines and an NVIDIA H100, that CPU-only use has not been tested, and that the full-precision weights take about 16 GB.