Nace.AI's Drex Skips Text Generation to Score Decisions in 183ms
Nace.AI releases four open decision models that return calibrated probabilities over options in milliseconds, skipping autoregressive decoding entirely.
Takeaways
- Nace.AI released Drex, four open decision models (8B, 9B, 35B MoE, 27B Max) that return option probabilities directly.
- Drex Max scores 67.52 on Decision Index 0.4.1, a new state of the art among decision models, beating Jev 1.14.0 at 60.70.
- Typical request latency is 183 ms versus about 2 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5.
- Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, plus consistency losses for option-order robustness.
- Weights, llama.cpp/Ollama/MLX runners and a reference server are on the Hugging Face org.
- Best for structured classification over documents; frontier reasoning models still lead on product-management and agent-evaluation tasks.
Drex, a family of open models from Nace.AI, accepts a document or agent state, a question, and a set of options. It returns a probability distribution over those options without generating prose. The approach targets routing, policy checks, document review, and other workflows where an application needs a structured choice rather than an open-ended answer.
The release includes four checkpoints documented in the research paper: the 8B DLM, the 9B v1.5, the 35B mixture-of-experts v2.0, and the 27B multimodal Max. The authors report that Max matches d3-turbo and GPT-6 Sol statistically on Decision Bench v4, trails Claude Opus 5.5 by 1.6 points, and answers a typical request in 183 milliseconds.
A classifier-shaped LLM
Many production systems ask a generative model to choose among known options, then parse its answer into an enum, Boolean, or score. Token-by-token generation adds latency and creates failure modes such as malformed JSON, unsupported labels, ambiguous prose, and conclusions that conflict with the accompanying explanation.
Drex scores the supplied options directly. Each question produces a fixed-shape probability distribution that an application can use for selection, thresholding, ranking, or abstention. This interface removes output parsing and retry logic, although teams still need to validate probability thresholds against their own data.
Four checkpoints, one request format
| Checkpoint | Size | Architecture or focus |
|---|---|---|
| DLM | 8B parameters | Original decision model |
| v1.5 | 9B parameters | Updated general checkpoint |
| v2.0 | 35B total, about 3B active | Mixture-of-experts model |