Distributed Reasoning Network ReasonNet Released: 1000 Small AI Models Collaboratively Outperform Single Large Models

Technical University of Berlin's Machine Learning Institute and the European AI League (EAIL) today jointly released ReasonNet, a novel distributed reasoning collaboration architecture. By dynamically decomposing complex reasoning tasks among up to 1000 specialized small AI models, ReasonNet has surpassed single large model approaches on multiple mathematical and logical reasoning benchmarks for the first time.

From Bigger is Better to Collaborative Division

Over the past three years, the AI industry's primary competitive direction has been building larger single models. But the ReasonNet team chose the opposite path from the start. Chief architect Katrin Bauer, Professor at TU Berlin, explained that human brain reasoning is not accomplished by a single super neuron but through collaboration among multiple specialized brain regions. ReasonNet mimics this pattern.

The system's core is a component called the Reasoning Dispatcher. When it receives a complex question, the dispatcher first analyzes the problem structure and required reasoning types, decomposes it into sub-problems, and assigns them to the most suitable small models. Each small model has between 7 and 13 billion parameters but can match or surpass general-purpose large models in its specialized domain.

Benchmark Results

On the GSM8K mathematical reasoning benchmark, ReasonNet achieved 97.3% accuracy, exceeding GPT-5's 95.8% and Gemini Ultra 2's 96.1%. On the more challenging MATH benchmark, ReasonNet scored 89.7%, beating the best single model by 4.2 percentage points. On logical reasoning (LogiQA 2.0), ReasonNet reached 91.2% versus the best single models 86.5%.

However, ReasonNet underperformed on open-domain QA tasks requiring broad world knowledge. Professor Bauer acknowledged that ReasonNet excels at structured problems requiring multi-step reasoning, while large models retain advantages for tasks requiring extensive knowledge retrieval.

Deployment Cost and Latency

A key advantage is deployment cost. The total compute demand of 1000 small models is approximately 30% of a trillion-parameter model, with inference latency matching single large models due to ReasonNet's parallel scheduling. TU Berlin has partnered with European cloud providers to launch a commercial version in Q3 2030.

The release could have profound implications for the industrys bigger is better paradigm. Nvidias Chief Scientist Bill Dally commented that ReasonNet represents a more democratized AI direction. However, analysts note that engineering complexity may limit adoption speed among smaller enterprises. The paper has been submitted to ICML 2030, with code and pretrained models to be open-sourced upon acceptance.