Deep Cogito v2: open-source models that internalize their own reasoning
San Francisco startup founded by ex-Googlers ships four open-source hybrid reasoning models — 70B, 109B, 405B, 671B — using a technique called Iterated Distillation and Amplification (IDA) to distill search-time reasoning back into model weights.
Deep Cogito's pitch is research-grade and unusual: rather than RLHF or pure inference-time chain-of-thought, the Cogito v2 family uses IDA — Iterated Distillation and Amplification — to take what the model figures out at search time and distill it back into the model itself. Net effect: each round of training, the model's intuition gets sharper because the discoveries from explicit search are absorbed.
The model sizes span 70B / 109B (mid-sized) and 405B / 671B (large). All released under permissive open-source licenses (MIT, Llama) — commercial use allowed, fine-tuned weights are yours, deploy anywhere.
Availability: Hugging Face for direct download, plus API via Together AI, Baseten, and RunPod. The 671B model puts Deep Cogito in the same scale bracket as DeepSeek's frontier checkpoints — uncommon for a US-based open-source startup.
This belongs on the watch list for anyone tracking whether the open-source frontier can keep pace with closed-frontier reasoning models.