Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text
Nace.AI has open-sourced Drex 1.5, a 9B decision model that returns a probability for every option in 1 forward pass. It scores 58.08 on Decision Index 0.3.1 and reads up to 128K tokens.
Nace.AI has open-sourced Drex 1.5, a 9B parameter decision model designed to evaluate options rather than generate text, with specific performance benchmarks and deployment flexibility.
What happened
Nace.AI released Drex 1.5, an open-source decision model that evaluates predefined options using a specialized architecture. The model is optimized for efficiency, with performance metrics and deployment options tailored for developers and enterprises.
Why it matters
Developers can leverage Drex 1.5 for applications requiring rapid decision-making over large datasets, though its limitations in general knowledge and sentiment analysis require careful use cases.
- Drex 1.5 offers open weights for transparency and customization
- Optimal for long-document decision tasks but not general-purpose AI
- Requires careful consideration of domain-specific limitations
- Multiple deployment options including cloud and local hardware
- Licensing restrictions apply under Nace.AI Open RAIL-M
The bigger picture
Provides an open alternative to closed decision models, enabling cost-effective deployment for specific tasks while highlighting the trade-offs between specialization and generalization in AI models.
Drex 1.5 represents a specialized open-source decision model with strong performance on specific benchmarks, offering developers a flexible tool for option-based decision systems while maintaining clear limitations in broader AI capabilities.
Technical details
- Built on MiMo-V2.6-Distill-Qwen-9B with hybrid attention mechanisms
- Pointer head architecture for option scoring from hidden states
- Supports 3 question types: choice, noul (yes/no), and ordinal score
- GGUF format for cross-platform deployment (Q8_0 variant for Apple silicon)
- Trained on index benchmark splits, evaluated on held-out data
FAQ
What distinguishes Drex 1.5 from traditional language models?
Drex 1.5 focuses on scoring predefined options rather than generating text, using a specialized architecture optimized for decision-making tasks.
How does Drex 1.5 handle long documents?
It achieves 93.4% accuracy on 32K-128K token documents, outperforming shorter token truncation results.
What are the main limitations of Drex 1.5?
Weaknesses in knowledge-intensive tasks (e.g., 45.4% GPQA Diamond accuracy) and fine-grained sentiment analysis (7.4% ACOS F1).
Can Drex 1.5 be deployed on consumer hardware?
Yes, the Q8_0 GGUF variant runs on Apple M5 Pro via Metal/CPU, and bf16 weights fit on a 24GB A10G GPU.
What licensing terms apply to Drex 1.5?
The model uses Nace.AI Open RAIL-M license with use restrictions, requiring review before commercial applications.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by MarkTechPost on Oct 10, 2026.
Comments
Sign in to join the discussion.