The AI startup Deep Cogito has released four new open source language models designed to get better at reasoning on their own. In an article for VentureBeat, Carl Franzen reported on the release from the San Francisco based company.
The models, part of the Cogito v2 family, are trained to internalize their own thought processes. According to the company, successful reasoning paths are distilled back into the models’ weights. This process helps the models develop what the company calls a “machine intuition”, allowing them to find correct answers more quickly and efficiently.
The four models range in size from 70 billion to 671 billion parameters and include both dense and Mixture-of-Experts (MoE) architectures. Deep Cogito claims its flagship 671B MoE model matches or exceeds the performance of other leading open models on reasoning benchmarks while using significantly shorter computational steps. For example, it reportedly solved a math problem using less than half the reasoning tokens of a competitor.
The company states it has achieved these results with high capital efficiency, training all its models for under $3.5 million. CEO Drishan Arora attributes this to a focus on smarter training rather than simply using more data. The models are available for developers on platforms like Hugging Face and through various API providers.