Nace AI Open-Sources Drex 1.5 Decision Model

Nace.AI has open-sourced Drex 1.5, a 9 billion parameter decision model specifically engineered for AI agents and backend workflows. Unlike traditional language models that generate text, Drex 1.5 operates by reading a given state and a set of typed questions, subsequently returning a probability score for each presented option. This distinct functionality positions it as a tool for decision-making processes rather than content creation. Nace reports that Drex 1.5 achieved a score of 58.08 on the public Decision Index 0.3.1 benchmark, which is noted as the top score among models with fewer than 10 billion parameters. The model's weights are made available on Hugging Face, and a hosted version is accessible through OpenRouter. Drex 1.5 features 8.95 billion dense parameters with bf16 weights, requiring approximately 18 GB of storage. Its default context window is set at 16,384 tokens, with the capability to extend up to 131,072 tokens. For deployment, the model can run on a single CUDA GPU with bf16 precision, having been tested on a 24 GB A10G card. A quantized Q8_0 GGUF version, approximately 9.5 GB, is also available and can operate on Apple silicon and CPUs. In terms of performance, Drex 1.5's score of 58.08 on the Decision Index 0.3.1 places it within the tie band of Jev 1.13.0, which scored 57.96. Nace highlights that the model demonstrates up to 93.4% accuracy on documents ranging from 32,000 to 128,000 tokens. However, its performance is weaker in specific areas, showing a 7.4% per-review F1 score on ACOS aspect sentiment, compared to Jev's 29.5%. The core advantage of Drex 1.5 lies in its open weights, which Nace claims match a closed-source model's performance while running on a single GPU. Conversely, its limitations include weaker performance on broad knowledge tasks and fine-grained sentiment analysis. Drex 1.5 is designed to score a fixed set of options in a single forward pass, receiving a state (in text or JSON format) and named questions. It supports three question types: 'choice', 'noul' (yes/no), and 'ordinal score'. Since no tokens are sampled, parameters like temperature and top_p are not applicable. The model is constrained to answer only with the options provided by the user. It serves requests via the POST /v1/systemone API, utilizing the same request format as TypeSafe's Jev, a closed model that pioneered this category. Nace states that existing Jev clients can be adapted to use Drex 1.5 by modifying a few environment variables. The underlying architecture of Drex 1.5 is based on MiMo-V2.6-Distill-Qwen-9B, a distilled version of the Qwen 3.5 9B model. This backbone comprises 32 layers featuring a hybrid attention mechanism, with three linear attention layers interspersed with each full attention layer. A distinct pointer head, identified as 'head.pt', is responsible for scoring each option by utilizing the backbone's hidden states. Each question processed by the model involves a single pass over the input state combined with that specific question. Within the llama.cpp framework, the state is encoded once and then shared across multiple questions, optimizing efficiency. Nace's dedicated Drex page indicates that the model was trained with a focus on decision-making capabilities.