Deep Cogito trained its Cogito v1 family on Runpod infrastructure on the road to ASI

San Francisco-based Deep Cogito trained its family of open Cogito v1 models on Runpod infrastructure. Runpod is a GPU cloud where customers rent compute instead of building their own cluster. The lineup comes in five sizes, from 3B to 70B parameters, and the company's stated goal is artificial superintelligence (ASI). The story was published on Runpod's blog as a customer case study, so it should be read with that genre in mind.

What we know about the project

There is only one public source here: a post on the Runpod blog. The excerpt we have cuts off at the line saying that Cogito v1 is aimed at artificial superintelligence. It does not disclose the number of GPUs, the accelerator types, the length of the training runs or the training cost. Without those figures, the scale of the compute cannot be assessed, and guessing at them would be wrong.

Deep Cogito's own public materials give a broader picture. The company emerged from stealth in April 2025 and introduced Cogito v1 Preview. The models were not trained from scratch: they are built on existing Llama and Qwen checkpoints, which received additional training using the company's own method.

How Cogito v1 works

Deep Cogito's main bet is IDA, iterated distillation and amplification. The scheme works like this. First, the model spends more compute on reasoning or search and arrives at a better answer than it would have given right away. That is the amplification step. Then the result is distilled back into the weights, and the model learns to produce that answer without the extra cost. The cycle repeats. The idea is that improvement depends not on human labeling but on the compute the team is willing to spend.

The models are hybrid: they answer directly or switch on a reasoning mode. The company has claimed that Cogito v1 outperforms open models of the same size, including the Llama, DeepSeek and Qwen lines. These are the company's own claims. Our article contains no independent verification.

Key parameters of the family:

• Sizes: 3B, 8B, 14B, 32B and 70B parameters.
• Base models: Llama and Qwen checkpoints.
• Modes: standard answer and reasoning (hybrid reasoning).
• Training method: iterated distillation and amplification (IDA).
• Infrastructure: Runpod cloud.
• Access: open weights.

Why the choice of Runpod matters

The case shows a typical path for a small lab. A cluster of thousands of H100s of one's own costs hundreds of millions of dollars and months of waiting for deliveries. Hourly rental lets a team start experiments right away and stop them when a method does not work. For IDA, where training is built on repeated cycles, that flexibility matters: the amount of compute needed for each iteration is not known in advance.

This is our assessment, not a quote from the source. For Runpod, which competes for customers with CoreWeave, Lambda and the hyperscalers, a case like this works as advertising. It tells the market that rented GPUs can be used to train models not only for applied tasks but also for bids at the frontier.

What to check in practice

A team considering Cogito v1 or a similar path should check a few things:

1. Run a model of the right size on your own tasks. The company's rankings are no substitute for a test on your data, especially in reasoning mode, where both latency and cost per token go up.

2. Read the licenses. The models are built on Llama and Qwen, so the terms of use are inherited from the base checkpoints.

3. If you plan your own fine-tuning on an open base, compare the economics of renting GPUs against buying them. For experiments with an unknown amount of compute, renting usually wins.

4. Watch whether Deep Cogito publishes technical details: the number of GPUs, the cost of the runs, data on IDA. Those are what will show whether the approach is reproducible outside this team.

For now, there is a claim, a working lineup of open models and a cloud provider that is happy to talk about it. That is not enough for a conversation about ASI, but as an example of a small team training competitive models without a data center of its own, the case works.