Deep Cogito Releases Cogito v2 Models, Challenging Frontier AI on a Budget

Deep Cogito has released Cogito v2, a suite of four open-weight reasoning models designed to deliver high performance without the computational bloat. The July 31 drop includes two dense models (70B and 405B) and two Mixture-of-Experts (MoE) variants (109B and 671B), all built using a novel training pipeline that fuses inference-time search with iterative self-improvement. The models are optimized to internalize reasoning strategies, reducing dependence on long, compute-intensive search chains.

The standout is the 671B MoE model, which the team says matches or outperforms DeepSeek R1 and v3, and approaches the quality of closed frontier models. Trained for less than $3.5 million, the release underscores how small teams can now produce cutting-edge reasoning systems on a fraction of traditional budgets. All models are open-licensed, extending practical access to developers and researchers aiming to build intelligent agents and reasoning-heavy applications.

Strategically, Cogito v2 positions Deep Cogito as a formidable player in the open model landscape. While OpenAI and Anthropic still dominate with larger, closed systems, Cogito v2 offers a competitive open-source path—especially for users prioritizing interpretability and integration flexibility. The project’s emphasis on reasoning also targets a key weakness in many open models, making this release a strong signal that reasoning quality may now be attainable without frontier-level compute.