Editorial illustration for Nace AI Open-Sources 9B-Parameter Drex 1.5 Decision Model
Nace AI Open-Sources Drex 1.5 Decision Model
Nace.AI released Drex 1.5 on Hugging Face this week, and the model skips the part most language models are built for. There's no text generation here. Feed it a state, in plain text or JSON, along with a set of named questions, and it returns a probability for each option you've defined. Nothing more.
The company built Drex 1.5 at 8.95 billion parameters, dense, with bf16 weights running around 18 GB. It handles 16,384 tokens of context by default and scales up to 131,072 when needed. A single CUDA GPU in bf16 will run it, Nace tested this on a 24 GB A10G, and a Q8_0 GGUF version at roughly 9.5 GB works on Apple silicon or plain CPU. The model answers three question formats: choice, yes/no, and ordinal score, all served through a POST /v1/systemone endpoint.
On the public Decision Index 0.3.1 benchmark, Nace reports a score of 58.08, the highest mark for any model under 10 billion parameters. That puts it inside the tie band of Jev 1.13.0, a closed model scoring 57.96. A hosted version is live now on OpenRouter alongside the open weights.
Nace.AI has open-sourced Drex 1.5, a 9B decision model for agents and backend workflows. The Drex 1.5 decision model does not write text. It reads a state and typed questions, then returns a probability for every option.
Why this matters
A 9B model that outputs probabilities instead of prose is a narrow bet, but it's the right kind of narrow. Most agent frameworks bolt decision logic onto a chat model and hope the text it generates maps cleanly onto an action. Drex 1.5 skips that translation step entirely: typed questions in, calibrated probabilities out. For developers building agents that need to pick between API calls, routing options, or next-steps in a workflow, that's a cleaner interface than parsing JSON out of an LLM's chain of thought.
The benchmark gap is thin, 58.08 versus 57.96 for Jev 1.13.0 isn't a blowout, and chance-corrected scores on 37 benchmarks deserve scrutiny before anyone treats this as settled. Training on the official splits of the index benchmarks is also worth flagging. It's not evidence of cheating, but it means third parties should run their own held-out evals before betting production systems on the ranking.
Still, open weights on Hugging Face and a live OpenRouter endpoint mean anyone can test this against their own decision workflows this week. That's the real signal: a usable, inspectable alternative to black-box routing logic, not just another leaderboard entry.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Anthropic AI Sent Fake Police Tip During Testing, Philly PD Says
From Archives OpenAI's Decisions API Enters Beta, Promises 10x Faster Typed Responses
Common Questions Answered
How does Drex 1.5 differ from traditional language models in terms of output?
Drex 1.5 does not generate text like traditional language models. Instead, it reads a state in plain text or JSON format along with named questions, then returns probability scores for each defined option. This focused approach eliminates the need to parse generated text into actionable decisions.
What are the technical specifications of the Drex 1.5 decision model?
Drex 1.5 is built with 8.95 billion parameters as a dense model using bf16 weights, requiring approximately 18 GB of storage. It handles 16,384 tokens of context by default and can scale up to 131,072 tokens when needed for larger decision contexts.
Why is Drex 1.5's probability-based output better for agent frameworks than text generation?
Drex 1.5 provides a cleaner interface for agents by outputting calibrated probabilities directly instead of requiring developers to parse generated text into actions. This eliminates the translation step between what a model writes and what action should actually be taken, making it ideal for routing API calls, choosing between options, or determining next steps in workflows.
What use cases is Drex 1.5 designed for?
Drex 1.5 is specifically designed for agents and backend workflows that need to make decisions between multiple options. It excels in scenarios where systems must choose between API calls, routing options, or determine the next step in a workflow based on the current state and available options.
Further Reading
- Introducing Drex, a small model that decides instead of writing - Nace.AI
- Drex, a small model that decides - Nace.AI
- nace-ai/drex-decision-models - Trendshift
- Nace.AI Drex v1.5 - OpenRouter
- nace-ai/drex-v1.5 - Hugging Face
