ansulev/Darwin-9B-NEG

Darwin-9B-NEG: The First Native Entropy Gating Model

Darwin-9B-NEG is a 9 billion parameter model based on a Qwen3.5-9B backbone, distinguished by its Native Entropy Gating (NEG) technology. NEG is a proprietary Darwin architectural innovation that integrates a "self-confidence" mechanism directly into the model's weights, allowing for self-regulated reasoning without external multi-turn iteration. This results in a significant +12.63 percentage point improvement in reasoning accuracy on the GPQA Diamond benchmark at the same 1x inference cost, with NEG activating in less than 5% of generation steps.

Key Capabilities & Features

• Native Entropy Gating (NEG): An internal mechanism that predicts next-token distribution entropy and guides token selection, enhancing reasoning accuracy without additional inference cost.
• High Reasoning Accuracy: Achieves 84.34% on GPQA Diamond (PhD-level reasoning) with a 3-stage ensemble protocol, surpassing the Qwen3.5-9B leaderboard result.
• Efficient Deployment: NEG is embedded within the model weights, requiring no extra libraries or complex setup; it loads directly with standard transformers.
• Evolutionary Model Merging: Part of the Darwin family, developed using Darwin V7, an evolutionary breeding engine that combines parent LLMs.

Recommended Use Cases

• Graduate-level STEM reasoning: Excels in physics, chemistry, biology, and mathematics.
• Mathematical problem solving: Suitable for MATH and AIME-style challenges.
• Code reasoning and debugging: Performs well on HumanEval-style tasks.
• Complex chain-of-thought tasks: Ideal for scenarios requiring enhanced reasoning from a smaller model.