Deep Cogito announced a $43 million Series A on August 26, 2026, led by TQ Ventures. Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler joined. The company said the round brings total funding to more than $56 million.

The San Francisco lab focuses on post-training: reinforcement learning and methods that teach a pre-trained model to reason and then improve on harder tasks. Founders Drishan Arora and Dhruv Malrana previously worked on Google’s AI Search products, including AI Mode and AI Overviews. Arora led Gemini post-training for AI Search. Malrana led the product from its start.

Open weights first, then enterprise models

Deep Cogito said it proved the methods on open-weight Cogito models from 3 billion to more than 600 billion parameters, then pointed the same engine at companies that want models trained on their own data and outcomes. Zscaler is both a customer and a strategic investor in the round.

One research line, Iterated Distillation and Amplification, lets a model spend extra compute to produce a better answer, then writes that gain back into the weights. The stated aim is models that improve themselves and, over time, move past the limits of human-written training data.

“Pre-training gives a model an enormous amount of knowledge and capability. Post-training determines what that model can actually become.” Drishan Arora, co-founder and CEO, Deep Cogito

Schuster Tanger, co-founding partner at TQ Ventures, said few teams outside the largest labs have shown they can post-train at this scale, and that Deep Cogito did it in public. The company said it will use the money to grow research and engineering, scale training infrastructure, ship later Cogito releases, and expand enterprise work.

Decoded Take

Post-training is where labs now spend the scarce part of the budget: the researchers who can turn a base model into something that holds up in a product. Deep Cogito is selling that layer twice, once as open weights that advertise the method, and once as a private engine for a customer’s own data. Zscaler in the cap table is the tell. A security company does not join a research round unless it already likes the specialized model it is running. The risk is that frontier labs keep the best post-training in-house and leave independents with last year’s base checkpoints. The watch item is the next Cogito drop. If a new open-weight model shows a clear jump from the same base that rivals already ship, the thesis holds. If the enterprise pitch stays private and the public models stall, this is a services shop with a research brand.