Cogito, a San Francisco-based, behavioral analytics startup, has very recently come out of stealth. They recently announced a new state-of-the-art set of models that embody the hybrid approach they pioneered. Cogito’s founders, former Google engineers Drishan Arora and Dhruv Malhotra, established Cogito in June 2024.
Weighting models
Their models are already starting to leap ahead in the increasingly competitive and chaotic world of artificial intelligence.
The centerpiece of Cogito’s debut is its flagship model, Cogito 70B, which have shown strong performance with accuracy, speed and precision. Remarkably, even with reasoning disabled, Cogito 70B still exceeds the performance of Meta’s freshly announced Llama 4 Scout model on the new LiveBench evaluation platform. This accomplishment further strengthens Cogito’s standing as one of the most powerful contenders in the race for AI language model supremacy.
A relatively small team created Cogito’s models within a span of approximately 75 days. To overcome this challenge, they took a unique route to enable the models to fluidly shift between reasoning and non-reasoning modes. This flexibility allows users to select the operational mode that most fits their needs. As opposed to purely reasoning-based architectures, this hybrid architecture unites reasoning components with standard, non-reasoning model types, expanding the capabilities of these models.
Cogito developed its technology on a well-established framework. They capitalized on technology from Meta’s open Llama and Alibaba’s Qwen models to develop a robust suite of creative solutions. Currently, our models span from 3 billion parameters all the way up to an industry-leading 70 billion parameters. Next, we hope to start rolling out models that fly all the way up to 671 billion parameters!
Internal benchmarking results indicate that Cogito’s largest model, the Cogito 70B, performs very well in reasoning capabilities. It does better than competitors such as DeepSeek’s R1 reasoning model on most mathematics and language assessments. This is a notable milestone, as Cogito asserts their models routinely outperform the best open models of similar size.
“Every model can either answer outright or think to themselves before answering (similar to reasoning models),” told a PR consultant for Deep Cogito. This dual capability underscores the flexibility of their models in addressing a wide range of queries.
Cogito’s founders were bullish about the future of their technology. We’re only beginning to come down the other side of our scaling curve. Up until now, we’ve barely tapped into the computing power typically used to train post-training large language models. This perspective embodies their dedication to ongoing improvement and scale.
They shared what they’re currently investigating in relation to complementary post-training methods of self-improvement. “Looking beyond the training we provide, we’re exploring proven, supplementary post-training methods to foster self-improvement,” the spokesperson shared.
Leave a Reply