FINAL-Bench/Darwin-9B-NEG

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 24, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

Darwin-9B-NEG is a 8.95 billion parameter model developed by FINAL-Bench, built on a Qwen3.5-9B backbone. It is the first model to feature Native Entropy Gating (NEG), a proprietary architectural innovation that embeds self-confidence directly into the model weights. This enables self-regulating reasoning at 1x inference cost, significantly boosting accuracy on complex tasks like graduate-level STEM reasoning and mathematical problem solving.

Loading preview...

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

Darwin-9B-NEG, developed by FINAL-Bench, is an 8.95 billion parameter model based on the Qwen3.5-9B architecture. It introduces Native Entropy Gating (NEG), a proprietary architectural innovation that integrates a "self-confidence sense" directly into the model's weights. This allows for self-regulating reasoning within a single decoding loop, activating in less than 5% of generation steps, and significantly improving reasoning accuracy without additional inference cost.

Key Capabilities & Differentiators

  • Native Entropy Gating (NEG): Achieves a +12.63 percentage point increase on the GPQA Diamond benchmark at 1x inference cost compared to its NEG-free baseline, Darwin-9B-Opus. This is facilitated by tiny learnable modules (NEG-Head and NEG-Gate) that predict and regulate the model's confidence.
  • High Reasoning Accuracy: Scores 84.34% on the GPQA Diamond PhD-level reasoning benchmark with its 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 engines for deployment, and is compatible with standard transformers loading.
  • Evolutionary Lineage: Part of the Darwin series, developed using the Darwin V7 evolutionary breeding engine, with Darwin-9B-Opus as its base.

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 when a small reasoning model with a significant boost is desired.

Finetunes
2 models