Overview
Drex 1.5 is Nace AI open-source 8.95B dense decision model. It reads a text or JSON state together with typed questions defined by the developer and returns a probability for every option without generating text, aimed at routing, selection and scoring inside agents.
Key Features
- Probabilities only: No token sampling: one forward pass yields a probability per option, so temperature and top_p no longer apply
- Three question types: Supports choice, yes/no and ordinal score over a POST /v1/systemone endpoint, and migrating from comparable closed-source options takes only a few environment variables
- Holds up on long documents: Reported accuracy is 93.4% on 32K to 128K token documents at 2.0 s median latency, versus 78.0% when the same requests are truncated to their first 8K tokens
- Runs on personal hardware: bf16 weights are about 18 GB and fit a single 24 GB GPU; a roughly 9.5 GB Q8_0 GGUF runs on Apple silicon and CPU
Use Cases
- Routing an agent between several APIs, with probability thresholds deciding which branch to take
- Scoring and grading long contracts or tickets, feeding the result straight into a downstream flow
- Replacing a text-parsing judgement step in a backend workflow, removing one format conversion
- Deploying the decision layer locally or on an internal network so business state never leaves your own infrastructure
Pros
- Calibrated probabilities come out directly, with no parsing step in between
- Highest published Decision Index score of any model under 10B parameters, inside the tie band of closed-source alternatives
- A GGUF quant is available, so personal devices and CPUs can run it
- The interface follows an existing format, so migration is a small change
Pricing
Weights are open on Hugging Face for self-hosting. The hosted version on OpenRouter is served by DeepInfra at $0.04 per million input tokens with free output. Context is 16,384 tokens by default and up to 131,072.
Summary
Drex 1.5 suits teams that need discrete judgements inside agents or backend workflows and want to control deployment themselves. Its value concentrates in long-document and multi-question scenarios; the real trade-off is whether a single-purpose inference runtime is worth maintaining.
Version History
- Nace AI open-sources the Drex 1.5 decision model (2026-10-10): Nace AI open-sourced Drex 1.5, an 8.95B dense decision model built on MiMo-V2.6-Distill-Qwen-9B. A pointer head scores each option directly from the backbone hidden states, and the state is encoded once and shared across questions. Context is 16,384 tokens by default and up to 131,072; bf16 weights are about 18 GB and run on a single 24 GB GPU, with a roughly 9.5 GB Q8_0 GGUF also available. Weights are on Hugging Face, and DeepInfra serves a hosted version on OpenRouter at $0.04 per million input tokens with free output.