nace-ai/drex-v2.0
Drex v2.0 by Nace.AI is a 35.1 billion parameter mixture-of-experts decision model built on Qwen3.6-35B-A3B, with approximately 3 billion active parameters per token and a 32,768 token context length, extendable to 131,072. It specializes in answering typed questions about a given context by providing probabilities for options, rather than generating text. This model excels in decision-making tasks, outperforming other models like Jev 1.13.0 on the public Decision Index 0.3.1 and in specific benchmarks like email classification and game theory scenarios.
Loading preview...
Drex v2.0: Nace.AI's Flagship Decision Model
Drex v2.0 is a specialized decision model developed by Nace.AI, designed to provide probabilistic answers to typed questions based on a given context. Unlike traditional generative LLMs, Drex v2.0 does not generate text; instead, it evaluates options and returns probabilities in a single forward pass. Built on the Qwen3.6-35B-A3B mixture-of-experts architecture, it features 35.1 billion parameters with about 3 billion active per token, and supports a context length of 32,768 tokens, extendable up to 131,072.
Key Capabilities
- Superior Decision-Making: Achieves a score of 61.25 on the public Decision Index 0.3.1, surpassing Jev 1.13.0 by 3.3 points and Drex v1.5 by 3.2 points.
- Contextual Question Answering: Processes a
state(context) and multiple named questions to return probabilities for each option. - High Performance in Specific Benchmarks: Demonstrates strong results in email classification (AUC 0.984 vs. Jev's 0.981) and competitive game theory scenarios, winning 60.9% of matches against Jev in an eight-game arena.
- Flexible Deployment: Runs efficiently on various hardware setups using Python, llama.cpp, or Ollama, with support for Apple Silicon (Metal) and CPU.
- Memory Efficiency: The Q8_0 GGUF quantization requires approximately 37 GB of memory, making it accessible for deployment on personal hardware.
Good for
- Automated Decision Systems: Ideal for applications requiring precise, probabilistic decision outputs from structured or semi-structured contexts.
- Classification Tasks: Particularly effective for tasks like email classification where high accuracy and recall at low false-positive rates are critical.
- Strategic Game AI: Suitable for developing AI agents that make informed decisions in complex game environments.
- Resource-Constrained Environments: Its optimized architecture and support for quantized models (Q8_0 GGUF) allow deployment on hardware with limited GPU memory, such as Apple Silicon devices.