Tonight the NEW MODELS channel on DEV·TV picked up Nace.AI: Drex v2.0, 11 minutes after it went live on OpenRouter.
The listing stood out because Drex doesn't chat. It decides.
What Drex v2.0 is
Drex v2.0 is a small decision model from Nace.AI. You send it a state (some context) plus typed questions, and it answers each one with a probability attached. There's no prose to parse.
It supports three question types:
- Yes/No: "Is this PR safe to auto-merge?"
- Multiple choice: "Which tool should the agent call next?"
- Score: "Rate this support ticket's urgency."
The OpenRouter listing:
| Context | 131K |
| Input | $0.06 / 1M tokens |
| Output | Free |
| API contract |
/v1/systemone (same as TypeSafe's Jev) |
The last row matters most for builders: code written against Jev's decision API can point at Drex with almost no changes. Decision models now have a shared interface, not just a shared idea.
Why "decision models" are suddenly a category
TypeSafe launched Jev on September 15, 2026 and called it a "System One" model, borrowing the term from Thinking, Fast and Slow. The pitch is that most LLM calls inside an agent aren't writing tasks at all. They're small judgment calls:
- route this request
- approve or reject this output
- score this item
- pick the next tool
Using a frontier model for these means paying for generated tokens and then parsing JSON back out of them, so you pay and add latency for prose nobody reads.
Other labs followed within about a month:
- Nace.AI launched Drex on Sep 24 and reported that it led the public Decision Index.
- Liquid AI's d1 scored 58.9 on Hugging Face's Decision Index 0.2.1, one point above Jev's 57.9.
- Microsoft-Decision-1 is queued up next on DEV·TV's NEW MODELS channel as I write this.
Under the hood: the open release
Alongside the hosted model, Nace ships an open, self-hostable version called Drex DLM:
- Backbone: NVIDIA Efficient-DLM-8B, a diffusion language model, with a decision adapter merged into its weights
- Pointer head: scores the answer options directly instead of decoding text, so you get calibrated probabilities over fixed choices in a single pass
-
Runners: a Python server, a llama.cpp fork, and an Ollama fork, all exposing
POST /v1/systemone - Agent skill: lets coding agents call a self-hosted server without a hosted API key
- License: the weights are CC BY-NC 4.0 (non-commercial) and the repo code is MIT
⚠️ Check the license before you ship anything commercial on the open weights. The hosted OpenRouter route is the simpler path for production.
Note that the architecture details above come from the open Drex DLM release. Nace hasn't said whether the hosted v2.0 shares the exact same architecture.
Where I'd actually use it
- Agent tool routing: "which tool next?" as a multiple-choice question
- Guardrails: a yes/no "safe to send?" check before every response
- Bulk triage: scoring or classifying thousands of rows
- Any
ifstatement you've been delegating to GPT or Claude
With free output and a few cents per million input tokens, you can afford to put a decision on every row of a dataset. That isn't realistic with a frontier model.
The probability also earns its place. A 0.97 can go straight through, while a 0.55 can fall back to a bigger model or a human. That gives you a built-in escalation path.
Who's behind it
Nace.AI is a Palo Alto startup. Walden Catalyst led its $21.5M seed round in May 2026, and it targets audit, compliance, and finance, where a wrong answer is expensive and a confidence score is a real requirement.
Catch them as they land
New models land daily, and the interesting ones are easy to miss. I built DEV·TV as a zero-infrastructure, TV-style feed for developer content. The NEW MODELS channel pulls fresh listings from OpenRouter, so releases like Drex show up on screen minutes after they're listed.
Have you tried a decision model in an agent pipeline yet? I'd like to hear where it worked and where it didn't. 👇
Top comments (0)