Deep Cogito
Deep Cogito builds open-weight Cogito language models and post-trains specialized models for enterprises.
Founded
2024Deep Cogito builds open-weight Cogito language models and post-trains specialized models for enterprises.
Deep Cogito publishes the Cogito family of open-weight large language models, each able to answer directly or engage a reasoning mode that reflects before responding. The releases are post-trained on top of open base models from Meta and Alibaba rather than pre-trained from scratch, and have spanned roughly three billion to 671 billion parameters.
The weights are published on Hugging Face and served through inference providers including Together AI, Fireworks AI, and Baseten. Alongside the open releases, Deep Cogito offers the same post-training engine to enterprises, training specialized models on a customer's proprietary data, decisions, and outcomes so that the customer owns the resulting model.
Deep Cogito is aimed at enterprise demand for models a company owns rather than rents, and its first customer points the same way: Zscaler engaged the lab before investing in it, and the two work together on security-specific models through Zscaler's Project AI-Guardian.
The company is part of NVIDIA's post-training partner ecosystem, received early access to the Nemotron models, and has said it will build Cogito v3 on them. It said the $43 million raised in August 2026 will expand the research and engineering team, scale training infrastructure, advance future Cogito releases, and grow the enterprise work.
Deep Cogito concentrates on post-training rather than pre-training, and its named research direction, Iterated Distillation and Amplification, spends extra computation to reach better answers and then trains those gains back into the weights so the improvement becomes permanent. Building on open base models let a small team ship its first hybrid reasoning family in about seventy-five days.
The open-weight releases serve as public proof that the method works at frontier scale, which the company converts into enterprise engagements. Its commercial argument is ownership: the customer keeps a specialized model trained on its own data rather than renting a general-purpose frontier model. Zscaler became both the first publicly named customer and a strategic investor.
Deep Cogito gives the Cogito models away as open weights, free to download from Hugging Face and to run on third-party inference clouds, so it charges nothing for the models themselves and takes no licensing or per-token revenue from them.
Revenue comes from enterprise post-training engagements instead. Deep Cogito works directly with a customer to understand its products and the metrics it measures, then trains that behavior into model weights the customer owns. The company publishes no pricing page and offers no self-serve signup.