We have noticed a very specific number from the Nace AI team: Drex Max scored 67.52 on the public Decision Index, 6.82 points above Jev 1.14.0, and takes roughly 183 milliseconds to handle a typical request. Specialized models are entering enterprise decision-making workflows with shorter response times.
What this is
Drex is a family of open models designed for “decision tasks”: given a document, record, or agent context—the information the system currently has—and a question with possible choices, the model directly assigns probabilities to each option instead of generating text step by step like a chat model. The four models use diffusion language models, hybrid attention, sparse mixture-of-experts, and dense architectures, with parameter sizes ranging from 8B to 35B.
The training data includes 1.15 million verified decision records and 2.9 million questions. Beyond supervised fine-tuning, the team added reinforcement learning and consistency training: the latter reduced the rate at which answers flipped when option order changed from 6.3% to 4.2%. Model weights are available in safetensors and GGUF formats and can run through llama.cpp and Ollama.
Industry view
The important shift is not “yet another model,” but the beginning of workflow decomposition: general-purpose chat models handle knowledge Q&A and content generation, while specialized models handle verifiable choices, reviews, and classification. Nace AI reports that Drex Max takes about 183 milliseconds per request, compared with roughly 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5. For high-concurrency businesses, that latency gap may matter commercially more than an overall leaderboard position.
The risks are equally clear. Nace AI acknowledges that these models still trail frontier chat models in closed-book knowledge and that the team has not yet completed strict controlled comparisons using the same backbone architecture. OpenRouter’s third-place ranking and Artificial Analysis’s top-five position mainly reflect performance on specific evaluations; they cannot be equated directly with results in real enterprise deployments. Nor does a decision model’s probabilistic output mean its judgments have been validated for the business. Once a wrong option enters risk control, hiring, or procurement, automation will simply amplify the error faster.
Impact on regular people
For enterprise IT: Businesses may gradually adopt combinations of “general-purpose models plus specialized models,” assigning tasks such as contract review and ticket triage to more stable, cacheable models while retaining human review and audit records.
For individual careers: More decision-making software will shift from “providing recommendations” to “directly recommending options.” Knowledge workers will need to pay closer attention to review conditions, probabilities, and exception rules, rather than focusing only on prompt-writing techniques.
For the consumer market: Local execution and lower latency will drive more lightweight assistants into customer service, shopping, and on-device applications, but faster performance does not automatically mean higher quality. Consumers will still need to verify data sources and determine who is responsible for outcomes.