Deep Cogito Raises $43M Series A to Build the Post-Training Engine for Self-Improving AI

Deep Cogito has raised a $43 million Series A as the San Francisco AI lab looks to scale an increasingly important part of the artificial intelligence stack: what happens after a foundation model has already been pre-trained.

The round was led by TQ Ventures, with participation from Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and cloud security company Zscaler, which is both a customer and strategic investor. The financing brings Deep Cogito’s total funding to more than $56 million.

Founded by Drishan Arora and Dhruv Malrana, who previously worked on Google’s AI Search products including AI Mode and AI Overviews, Deep Cogito is betting that future gains in AI intelligence will increasingly come from better learning algorithms and post-training rather than simply making pre-training runs larger.

That puts the company in a different part of the AI race from labs primarily focused on building ever-larger foundation models from scratch.

Why Deep Cogito is betting on post-training

Pre-training is responsible for giving a large language model its broad base of knowledge by exposing it to enormous quantities of data. Post-training then shapes how that model behaves, reasons, follows instructions, uses tools, and performs specific tasks.

Deep Cogito’s thesis is that this second stage has considerably more room to evolve.

The company’s research concentrates on large-scale reinforcement learning and methods that allow models to progressively improve their own capabilities. Rather than treating a pre-trained model as a mostly finished product that needs some additional alignment, Deep Cogito views it as a starting point whose underlying capabilities can continue to be developed.

Arora summarized the distinction in announcing the round: “Pre-training gives a model an enormous amount of knowledge and capability. Post-training determines what that model can actually become.”

This is particularly relevant as the economics of simply scaling pre-training become more demanding. Training frontier foundation models requires enormous datasets, GPU clusters, power infrastructure, and capital. A sufficiently effective post-training system could potentially extract substantially more intelligence from an existing model without repeating that entire process.

Turning expensive reasoning into better intuition

At the center of Deep Cogito’s research is Iterated Distillation and Amplification (IDA), an approach the company introduced publicly with its first Cogito model family.

The concept is relatively straightforward even if implementing it at scale is not.

During the amplification stage, a model is given additional computation and techniques that allow it to arrive at a stronger answer than it might generate immediately. The resulting improvement is then distilled back into the model’s parameters. That upgraded model becomes the starting point for another iteration.

Instead of requiring the model to perform an increasingly long reasoning process every time it encounters a difficult question, the objective is to gradually internalize some of what it learned during those more computationally expensive reasoning steps.

Deep Cogito describes this as improving a model’s “intuition.”

The distinction could become important for inference economics. Long reasoning chains can improve accuracy, but they also increase token consumption, latency, and cost. A model that has internalized better reasoning trajectories may be able to reach similar conclusions with less inference-time computation.

Cogito models provide a public test of the approach

Deep Cogito has been using its Cogito family of open-weight models as a testing ground for these techniques.

Its initial releases ranged from 3 billion to 70 billion parameters, before the company expanded the family with 70B, 109B mixture-of-experts (MoE), 405B, and 671B MoE models.

With Cogito v2, the company reported that its 671B model produced reasoning chains roughly 60% shorter than DeepSeek R1 0528 while remaining competitive across several evaluations. Deep Cogito also said that it spent less than $3.5 million combined training eight Cogito models ranging from 3B to 671B, although benchmark performance and training-cost comparisons do not necessarily translate directly into production economics.

The company continued refining the system with Cogito v2.1 671B, which uses an open-licensed DeepSeek base model that Deep Cogito then post-trained in-house. Its v2.1 research release says the model uses process supervision during reasoning, with training designed to improve the model’s ability to identify productive reasoning paths rather than simply encouraging it to reason for longer.

This progression is important to Deep Cogito’s broader argument. The company does not need to prove that it can pre-train a better foundation model than the largest AI labs. It needs to demonstrate that its post-training process can consistently turn strong existing models into more capable ones.

Zscaler points to the enterprise opportunity

The other side of the business is taking those post-training techniques beyond publicly released Cogito models and applying them to enterprise AI.

Instead of relying exclusively on a general-purpose frontier model, companies can potentially train specialized models around their proprietary data, workflows, evaluation criteria, and desired outcomes.

That can go considerably deeper than retrieval-augmented generation (RAG), which typically gives a general model access to external company information at inference time. Deep Cogito’s approach aims to change the model itself by training domain-specific capabilities into its weights.

Zscaler provides an early example. The cybersecurity company began as a Deep Cogito customer before participating in the Series A. The companies have also worked together around specialized security intelligence, with Deep Cogito arguing that security applications can benefit from models post-trained on an organization’s own security data and outcomes.

The broader enterprise proposition is about both specialization and control. The Wall Street Journal’s coverage of the financing notes that Deep Cogito is positioning its technology around models that businesses can train and own using proprietary data, rather than remaining entirely dependent on closed frontier systems.

What the $43M Series A will fund

Deep Cogito plans to use the Series A to expand its research and engineering organization, increase the computing infrastructure available for large-scale training, develop future Cogito models, and work with more enterprises seeking specialized models.

Infrastructure will likely be a particularly important part of that expansion. Deep Cogito’s own hiring materials describe researchers working across novel reinforcement learning algorithms, data pipelines, evaluation, and distributed infrastructure while training models exceeding 400 billion parameters.

The financing therefore gives the company considerably more capacity to test whether the improvements demonstrated through its early models continue as it applies more computation and runs more iterations of its post-training process.

The bigger implication: AI that learns from its own reasoning

The larger significance of Deep Cogito’s work is the possibility that AI development could shift from systems trained mostly on human-generated data toward models that increasingly generate, evaluate, and internalize their own improvements.

Approaches such as Iterated Distillation and Amplification aim to turn expensive reasoning into lasting capability. A model can use additional computation to solve harder problems, then distill those gains back into its weights, creating a stronger starting point for the next training cycle.

Over time, this could produce a development loop of reason, evaluate, learn, and improve.

There is still a major gap between iterative post-training and true recursive self-improvement. Models need reliable evaluations, carefully designed objectives, and safeguards against reinforcing mistakes. But even limited versions of this approach could reduce dependence on ever-larger pre-training runs and make specialized AI trained around proprietary enterprise data more practical.

If that trajectory continues, competitive advantage in AI may depend less exclusively on who has the largest compute clusters and more on whose models are best at learning from their own experience.