• Deep Cogito emerges from stealth with $20M seed funding to build hybrid AI models.
  • The startup fuses symbolic reasoning with deep learning to boost AI's ability to think and plan.
  • It aims to bridge the gap between generative AI and true machine reasoning.

A new AI research startup, Deep Cogito, based in San Francisco, has officially emerged from stealth with Cogito VI.

Fine-tuned from Meta's Llama 3.2 and equipped with hybrid reasoning capabilities. Cogito VI is a new line of open source large language model (LLM), with the ability to answer quickly and immediately, or "self-reflect".

Most of the reasoning models, like OpenAI's o1 and DeepSeek R1, can effectively fact-check themselves by working through complex problems step by step.

Deep Cogito seeks to expand the capabilities of AI beyond the existing constraints of human oversight by allowing models to continuously enhance and internalize their own advanced reasoning techniques.

As most hybrid models, Cogito can quickly answer queries while spending additional time considering more challenging queries.

Cogito claims that they outperform the best open models of the same size, including models from Meta and Chinese AI startup DeepSeek.

“Each model can answer directly […] or self-reflect before answering (like reasoning models),” the company explained in a blog post. “[All] were developed by a small team in approximately 75 days.”

They are offered under the Llama licensing terms, which permit commercial use, thereby allowing third-party companies to utilize them in paid products.

This is applicable for up to 700 million monthly users; beyond that threshold, a paid license from Meta is required.

The company also plans to release larger models, including 109B, 400B, 671B, in the coming weeks, as well as improved checkpoints for each of these model sizes.

The Cogito 1 models vary in size from 3 billion to 70 billion parameters, and Cogito has announced that models with up to 671 billion parameters will be added in the upcoming weeks and months.

Parameters are a rough indicator of a model's problem-solving abilities, with a higher number typically indicating better performance.

Based on Cogito's internal benchmarking results, the largest model, Cogito 70B, demonstrates superior reasoning capabilities compared to DeepSeek's R1 reasoning model in several mathematics and language assessments.

Additionally, when reasoning is disabled, Cogito 70B also outperforms Meta's recently launched Llama 4 Scout model on LiveBench, a general-purpose AI evaluation.

“Currently, we’re still in the early stages of [our] scaling curve, having used only a fraction of compute typically reserved for traditional large language model post/continued training,” Cogito wrote in its blog post.
“Moving forward, we’re investigating complementary post-training approaches for self-improvement.”

Although the company is still in the early stages, it is aiming to build reasoning systems that could eventually power AI agents, enterprise copilots and autonomous decision-makers.

Edited by Annette George