Drex v2.0 Released: A Decision Model That Makes Judgments Instead of Writing Prose

Does an AI model have to write essays to be considered intelligent? Nace.AI offers a different answer with Drex v2.0: it doesn't chat, it only makes judgments.

DevTV's NEW MODELS channel picked up Drex v2.0 just 11 minutes after it went live on OpenRouter. That speed alone says something — decision models are becoming a distinct category, not just a feature bolted onto some large model.

What Drex v2.0 Actually Does

Drex v2.0 is a small decision model. You give it a chunk of state (context), attach a set of typed questions, and it returns answers with probabilities, one by one. No prose, no natural language to parse.

It supports three question types:

• Yes/no questions: Can this PR be auto-merged?
• Multiple choice: Which tool should the agent call next?
• Scoring: Rate the urgency of this customer support ticket.

The pricing info on OpenRouter is also interesting: 131K context window, $0.06 / 1M tokens for input, free for output. Output is free because it doesn't generate tokens at all — the model directly scores candidate answers through a pointer head, producing a calibrated probability distribution in a single forward pass.

The thing builders should care about most is the API contract: /v1/systemone, identical to TypeSafe's Jev. That means decision-calling code written for Jev can switch to Drex with almost no changes.

Diffusion Language Models (DLMs) are a class of text models that borrow the diffusion mechanism from image generation. Traditional autoregressive models generate token by token, left to right, while diffusion language models first initialize the entire sequence as noise, then restore the text through multiple denoising steps. The advantage of this architecture is that it can process the whole sequence in parallel and naturally supports scoring over a fixed candidate set — because the model can evaluate all options simultaneously within the same denoising step. NVIDIA's Efficient-DLM series is a backbone network designed precisely for this kind of parallel, non-autoregressive scenario. After Drex DLM merges the decision adapter into the weights, the model no longer needs to generate any text, but directly uses the hidden states output by the diffusion process to probabilistically evaluate candidate answers — this is the fundamental reason for its zero token cost on the output side.

Pointer Head is a common neural network output layer structure, first popularized in sequence-to-sequence tasks like pointer-generator networks. Unlike a traditional classification head, a pointer head doesn't predict a distribution over a fixed vocabulary, but instead assigns probability weights to candidate positions or candidate sets already present in the input. In the decision model scenario, the pointer head receives the model's encoded representation of the state, then directly scores each of the several typed question options provided by the caller, outputting normalized probabilities. Since the candidate answers are finite and pre-given, the model doesn't need to decode text token by token — a single forward pass yields the calibrated probability distribution for all answers. This both avoids the engineering burden of JSON parsing and pushes inference cost to extremely low levels, making it a key engineering choice that distinguishes decision models from general-purpose large models.

Why Decision Models Suddenly Became a Category

TypeSafe released Jev on September 15, 2026, defining it for the first time as a "System One" model, borrowing the concept from Thinking, Fast and Slow. Its core insight: most LLM calls inside an agent aren't writing tasks — they're small judgments about routing requests, approving or rejecting outputs, scoring items, and picking the next tool.

Using frontier large models for these tasks means you pay for the tokens generated and also have to parse JSON out of prose. What you're paying for is text nobody will ever read, plus extra latency.

Within a month, multiple labs followed suit. Nace.AI launched Drex on September 24, claiming the top spot on the public decision index. Liquid AI's d1 scored 58.9 on the Hugging Face Decision Index 0.2.1, one point higher than Jev's 57.9. On DevTV's NEW MODELS channel at the time of writing, Microsoft-Decision-1 was already queued for its debut.

Multiple independent products emerging within the same time window, all adopting similar interface designs, shows that "decision" has gone from a vague idea to a consensus engineering problem.

System One and System Two are concepts from psychologist Daniel Kahneman's book Thinking, Fast and Slow. System One refers to fast, intuitive, automatic cognitive processes that consume almost no attention; System Two is the slow, deliberate reasoning process that requires working memory. TypeSafe borrows this pair of concepts to distinguish two kinds of model calls inside an agent: tasks requiring complex text generation and multi-step reasoning go to frontier large models (System Two), while high-frequency, low-complexity judgments — routing, approval, scoring, tool selection — go to dedicated System One decision models. The significance of this naming is that it clearly separates "decision" from "generation," carving out an independent and reasonable ecological niche for specialized small models, and making it easier for developers to think at the architecture level about which calls should go down which path.

Technical Details of the Open-Source Drex DLM

In addition to the hosted model, Nace also released a self-hostable version, Drex DLM:

• Backbone: NVIDIA Efficient-DLM-8B, a diffusion language model, with the decision adapter merged into the weights
• Pointer head: Directly scores answer options rather than decoding text, yielding calibrated probabilities over fixed options in a single forward pass
• Runners: Python server, llama.cpp fork, Ollama fork, all exposing POST /v1/systemone
• Agent skill: Lets coding agents call the self-hosted service without a hosted API key
• License: Weights under CC BY-NC 4.0 (non-commercial), repository code under MIT

Note that the above architecture details come from the open-source Drex DLM version. Nace has not yet said whether hosted v2.0 uses the exact same architecture.

If you plan to use the open-source weights in a commercial product, be sure to check the license terms first. For production, going the OpenRouter hosted route is less hassle.

CC BY-NC 4.0 is the Attribution-NonCommercial 4.0 International license published by Creative Commons. It allows others to copy, distribute, and adapt the work provided the original author is credited, but explicitly prohibits using the work for commercial purposes. For model weights, this means you're free to download, study, and fine-tune Drex DLM and use it in personal projects or academic research, but you cannot integrate it into customer-facing paid products, SaaS services, or internal commercial workflows. By contrast, the repository code uses the very permissive MIT license, which allows commercial use, modification, and redistribution, requiring only that the copyright notice be retained. This "non-commercial weights + commercial code" combination is not uncommon in the open-source model community; its purpose is to protect the model's commercial value while lowering the barrier to engineering integration. To deploy in a commercial product, you should prefer the OpenRouter hosted version or contact Nace for a commercial license.

Real-World Use Cases

The value of decision models isn't in generality, but in separating "judgment" from "generation." A few high-value scenarios:

• Agent tool routing: "Which tool to use next" as a multiple-choice question
• Safety guardrails: A yes/no check on "is this safe to send" before every reply
• Batch triage: Scoring or classifying thousands of data rows
• Any if statement you used to hand off to GPT or Claude

With free output and input at just a few cents per million tokens, you can run a decision on every single row of a dataset — something completely impractical with frontier models.

Probabilistic output also comes with a built-in escalation path: 0.97 confidence can pass straight through, while 0.55 can fall back to a larger model or human review. This tiered handling mechanism is especially valuable in compliance and audit scenarios.

The Team Behind It

Nace.AI is a startup based in Palo Alto that closed a $21.5 million seed round led by Walden Catalyst in May 2026. Its target markets are audit, compliance, and finance — fields where wrong answers are expensive and confidence scores are a must-have.

What to Watch

The rise of decision models reflects a pragmatic shift in agent engineering: not every step needs a model that writes essays. Handing judgment-type calls to specialized small models cuts costs, reduces latency, and naturally provides interpretable probabilistic output.

The fact that Drex v2.0 shares an API contract with Jev may matter more than the model's own performance. When a category has a unified interface, that's when the ecosystem truly starts to grow.

Have you tried decision models in your agent pipelines? Share which scenarios worked well and which didn't.