The Story
A model that thinks about how it thinks
Filed under: things that sound recursive because they are
Here is a reasonable thing to be skeptical about. A small AI company comes out of stealth, says it is building "general superintelligence," and releases some open-source models. Everyone says superintelligence now. It is the sort of claim that costs nothing to make and, historically, has cost quite a lot to keep. So the interesting question about Deep Cogito is not the mission statement, which is standard-issue Bay Area cosmic ambition. The interesting question is the method - because the method is specific, it is testable, and it is a little strange.
Deep Cogito was founded in 2024 in San Francisco by Drishan Arora, who had been a senior software engineer at Google leading language-model work for its generative search product, and Dhruv Malhotra, a former product manager at Google DeepMind. In April 2025 the company left stealth with Cogito v1: a family of open models, fine-tuned from Meta's Llama and Alibaba's Qwen, that shipped with a feature the company calls hybrid reasoning. The pitch is simple enough to fit on an index card. The same model can answer you directly, the way a normal chatbot does, or it can stop and reason through the problem first, the way the expensive "reasoning" models do. One model, two modes, and you pick per query.
That is a nice product idea, and it is not why anyone should pay attention. The reason to pay attention is a training method with a name that sounds like a physics lecture: Iterated Distillation and Amplification, or IDA. It is worth slowing down on, because it is the whole company.
We are building general superintelligence - not only to match human-level abilities but also to uncover entirely new capabilities we have yet to imagine.
— Deep Cogito, company mission
Most model training works one of two ways. Either you pay humans to rate answers and you nudge the model toward the highly-rated ones, or you take a big, smart model and use it as a teacher to train a smaller, cheaper student. Both approaches share a ceiling: the student can only get as good as its teacher, and the teacher is either a bigger model or a room full of human labelers. You are, in a sense, always borrowing intelligence from somewhere else.
IDA tries to remove the lender. In the amplification step, the model spends extra compute reasoning hard about a problem - thinking longer, exploring more paths - and arrives at a better answer than it could produce off the cuff. In the distillation step, that better answer gets folded back into the model's own weights, so that next time the same quality comes out faster and cheaper. Then you repeat. The model teaches itself, amplifies, distills, and repeats, and the claim is that intelligence can scale through compute and algorithm design rather than through human labels or a larger teacher standing behind it.
How IDA works
The self-improvement loop
STEP 01
Amplify
The model spends extra compute reasoning harder, reaching a better answer than its instant one.
→
STEP 02
Distill
That better answer is folded back into the model's own weights - no human labels.
→
STEP 03
Iterate
Repeat. Each pass sharpens the model's intuition for which reasoning path to trust.
↻ the output of step 3 becomes the input of step 1 ↺
If that sounds too clean to be real, the useful counterweight is the evidence Deep Cogito put on the table. The company reports it trained its initial model family with a small team in roughly 75 days - a genuinely short cycle for competitive models, and the kind of number that only looks good if the method is doing real work. When the Cogito v1 models shipped, they climbed open-model leaderboards quickly. The company's own benchmarking claimed its 70B model with reasoning outperformed DeepSeek's R1 on select math and language evaluations, and beat Meta's Llama 4 Scout on LiveBench without reasoning at all.
Then, in August 2025, Deep Cogito went bigger. The Cogito v2 family spans 70B, a 109B mixture-of-experts model, 405B, and a flagship 671B mixture-of-experts model. The headline is that the 671B open model lands among the strongest open models in the world - matching or exceeding DeepSeek's v3 and R1 and approaching closed frontier systems like OpenAI's o3 and Anthropic's Claude Opus. But the detail that actually captures the philosophy is smaller and better: Cogito's 671B reportedly reaches its answers using roughly 60% shorter reasoning chains than DeepSeek R1.
Each model can answer directly, or self-reflect before answering, like reasoning models.
— On Cogito's hybrid design
That 60% is the tell. A lot of the industry treats reasoning as something you buy by the yard - longer chains of thought, more tokens, more compute, a bigger bill. Deep Cogito's argument is that the win is not longer thinking but better-aimed thinking: teaching the model which reasoning path to trust so it stops meandering. In their framing, they give the model a signal about the quality of the thinking process itself, so it develops intuition for the right trajectory instead of wandering toward the answer. Shorter chains are not a compromise. For anyone paying for inference, they are the product.
The name, for what it is worth, is a small joke that improves the more you know. "Cogito" is Descartes - cogito, ergo sum, I think therefore I am - which is a fitting label for a model whose entire training story is about thinking about its own thinking. The logo is an icosahedron, twenty triangular faces, complex but structured, which is either a statement about intelligence or just a nice shape. Probably both.
