Deep Cogito v2: Open-source AI that hones its reasoning skills

Deep Cogito has released Cogito v2, a new family of open-source AI models that sharpen their own reasoning skills.

Released under an open-source licence, the new Cogito v2 lineup includes four hybrid reasoning AI models: two mid-sized at 70B and 109B parameters, and two large-scale versions at 405B and 671B. 

The largest, a 671B Mixture-of-Experts (MoE) model, is already being touted as one of the most powerful open-source AIs in the world. The company reports that it competes with the latest from DeepSeek and is closing the gap on proprietary systems like O3 and Claude 4 Opus.

But the real story isn’t just about size or power; it’s about a fundamental shift in how the AI learns. Instead of just ‘thinking’ longer at inference time to find an answer, Cogito v2 is designed to internalise its own reasoning processes.

This internalised reasoning is achieved through a technique called Iterated Distillation and Amplification (IDA), which distils the discoveries from a search back into the model’s core parameters. The goal is to build a stronger ‘intuition’, allowing the model to anticipate the outcome of its own reasoning without having to perform the entire search.

Because the open-source AI models have a better “gut feeling” for the right approach, their reasoning chains are 60% shorter than those of rivals like Deepseek R1.

This efficiency extends to the budget. Deep Cogito says that it developed all its models – from experiments to final training – for a combined total of less than $3.5 million. Still a large sum likely for you or I, but miniscule compared to the spending of many of the leading AI labs.

The flagship 671B model received special attention, trained not only to improve its final answers but to refine the thinking process itself. This approach discourages the model from “meandering” and rewards a more direct path to the solution. The performance data suggests it works, with Deep Cogito’s open-source AI model matching or exceeding the latest DeepSeek versions on key benchmarks while being close to proprietary alternatives:

Benchmark comparison of the flagship open-source Deep Cogito v2 671B AI reasoning model against Deepseek and OpenAI o3 and Anthropic Claude models.

Perhaps one of the most surprising outcomes is the models’ ability to reason about images; a skill they were never explicitly trained for.

The team shared an example of this reasoning where Deep Cogito’s open-source AI model compared two images of a duck and a lion, demonstrating a deep thinking process about their habitats, colours, and composition purely through transfer learning. Deep Cogito believes this emergent property could be a powerful way to bootstrap training data for future multimodal reasoning systems.

Looking ahead, the Deep Cogito team plans to “hill climb on the gains of iterative self-improvement” in its quest to build superintelligence. They have restated their commitment that all AI models they create will be open-source.

Source: https://www.artificialintelligence-news.com/