Hugging Face aquif-ai/aquif-3.6-8B is an 8 billion parameter hybrid reasoning model developed by aquif-ai, featuring a dynamic thinking mechanism that automatically adjusts reasoning depth based on query complexity. This model achieves a 4% performance improvement and 28% better token efficiency compared to its predecessor, aquif-3.5-8B-Think. It is optimized for tasks requiring intelligent resource allocation and efficient problem-solving, particularly in areas like mathematical reasoning and code generation.
Overview
aquif-3.6-8B is an 8 billion parameter hybrid reasoning model from aquif-ai, designed to dynamically determine the necessity and depth of its reasoning process based on the complexity of the input query. Building upon aquif-3.5-8B-Think and leveraging AutoThink RL data, this model intelligently allocates cognitive effort, similar to KAT-V1's approach but with autonomous control.
Key Capabilities
- Dynamic Reasoning: Automatically decides when and how deeply to engage in reasoning, adapting to task complexity.
- Enhanced Efficiency: Achieves a 28% reduction in token usage and a 12% decrease in thinking frequency on average, leading to more cost-effective operations.
- Improved Performance: Demonstrates a 4% average performance improvement across various benchmarks, including AIME 2025, LiveCodeBench, and GPQA Diamond.
- Smart Resource Allocation: Optimizes computational resources by only thinking when necessary, unlike models requiring manual reasoning toggles.
Performance Highlights
- AIME 2025: +1% performance with 26% fewer tokens.
- LiveCodeBench: +4% performance with 32% fewer tokens.
- GPQA Diamond: +6% performance with 24% fewer tokens.
Good For
- Applications requiring efficient and intelligent problem-solving.
- Scenarios where dynamic resource allocation for reasoning can lead to cost savings and faster inference.
- Tasks benefiting from improved performance in mathematical reasoning and code-related challenges.