# Drex open decision models return probability distributions over options without autoregressive decoding

> Drex Max, a 27B-parameter open-source decision model, scores 67.52 on Decision Index 0.4.1 and answers in 183 milliseconds-roughly 11 times faster than GPT-6 Sol.

- Type: AI news
- Page: https://completeaitraining.com/news/drex-open-decision-models-return-probability-distributions/
- Published: 2026-11-06
- Updated: 2026-11-06
- Author: Jeroen Erne
- Jobs: Education, Insurance, IT and development, Management, Product development
- Reading time: 2 min

Nace AI released Drex, a family of four open-source decision models that return probability distributions over user-supplied options without autoregressive decoding. The models range from 8 billion to 27 billion parameters, and the largest - Drex Max - sets a new accuracy record among comparable decision models while answering typical requests in 183 milliseconds.

#### How decision models differ from autoregressive approaches

Large language models using long chain-of-thought reasoning have become standard for AI applications because they achieve higher accuracy across many tasks. But in production, the authors note, they "give inconsistent answers, require parsing of outputs, and add cost and latency." Decision models take a different path. A user provides a question and a set of options, and the model returns a probability distribution directly - no text generation, no parsing, no decoding step.

Each request to Drex includes a state - a document, record, or agent context, optionally with images - and questions about it. Questions can ask for a choice among options, a yes/no judgment, or a score on an ordered scale. This structure suits workflows in insurance underwriting, product configuration, compliance checks, and other enterprise settings where answers must be structured and repeatable.

#### Training with supervised fine-tuning, RLAF, and consistency losses

The Drex family includes four models: an 8B parameter model (DLM), a 9B model (v1.5), a 35B model with 3B active parameters (v2.0), and the 27B Max model. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, or RLAF. Vision tuning was added for Max, giving it the ability to process images alongside text in the state field.

Ablation studies showed two clear patterns. RLAF improved accuracy across the board. Consistency losses - penalties applied when the model's output shifted due to option reordering - improved robustness. The result is a model whose decisions stay stable even when the order of choices changes, a known pain point for autoregressive systems.

#### Benchmark results and latency comparisons

On Decision Index 0.4.1, Drex Max scored 67.52 on the public suite, compared to 60.70 for Jev 1.14.0. On Decision Bench v4, Max was statistically tied with d3-turbo and GPT-6 Sol, and trailed Claude Opus 5.5 by 1.6 points. It also posted the highest accuracy among decision models on DecideBench.

Latency tells a sharper story. Max answers a typical request in 183 milliseconds. GPT-6 Sol takes about 2.0 seconds, and Claude Opus 5.5 takes about 3.5 seconds. For production systems handling thousands of decisions per minute, that gap compounds quickly. All four models are available under an open-source license on GitHub, making them accessible for teams building [AI engineering](https://completeaitraining.com/learningpath/ai-for-software-engineers/) pipelines that need structured, low-latency decision outputs.

#### Why this matters for IT, product, and insurance professionals

For teams building automated decision pipelines - claims triage in insurance, eligibility checks in benefits administration, feature flagging in product development - Drex offers a concrete alternative to general-purpose LLMs. The combination of structured output, sub-200ms latency, and open-source licensing removes three common barriers to production deployment: inconsistent formatting, slow response times, and vendor lock-in. The consistency guarantees around option ordering also reduce a class of errors that can surface in compliance audits or customer-facing decisions.

---

This is the Markdown form of https://completeaitraining.com/news/drex-open-decision-models-return-probability-distributions/. The whole site is mapped at https://completeaitraining.com/llms.txt.