Deep Cogito Companies need AI models that understand how their teams make decisions. A general-purpose model starts with broad knowledge. Companies still have to teach it their products, customers, code, security policies, and internal standards. RAG can retrieve a relevant document at inference time. Specialized work requires the model to learn from a company’s examples and outcomes. @DeepCogito is solving this through post-training, using reinforcement learning to train open-weight models on company-specific work. Deep Cogito also researches recursive self-improvement: models use additional compute to produce better answers, then learn from those answers through further training. Over time, those improvements become part of the model itself. This gives companies more control over their AI. They can build models around their own data, reduce dependence on closed model providers, and manage the cost of running AI at scale. It's been awesome to see Deep Cogito couple their ambitious research agenda with real execution. The team has released its Cogito family of open-weight models, and is running large-scale post-training programs with technology companies such as Zscaler. Building at the frontier of AI research is hard. Turning that research into models that companies can use is even harder. Deep Cogito is doing both, and doing both well. I’m excited by the scale of @drishanarora and @dhruvmalrana's ambition. Their long-term goal is to build models that keep improving beyond the limits of human-generated training data. Their work will also strengthen the American open-weight model community at a time when that capability matters enormously. Congrats to Drishan, Dhruv and the @DeepCogito team on their $43M Series A. This is a great milestone for a team pursuing one of the hardest and most consequential problems in AI.

Most tech startups in the US and elsewhere use Chinese open source models. Today, tech companies have a best-in-class US alternative. Congrats to @drishanarora Arora and DeepCogito team for releasing the best open-weight LLM by a US company: Cogito v2.1 671B. The model performs better than every US open model (GPT-OSS, Llama, Nemotron) on industry benchmarks. Importantly, it uses significantly fewer tokens because it has better reasoning capabilities, AND delivers improvements across instruction following, coding, longer queries, multi-turn and creativity. Try it out in one of the following ways: 1. check it out at the free interface below 2. download it on Hugging Face 3. try it out on OpenRouter, Inc, Together AI, Fireworks AI, Ollama cloud, Runpod, Baseten 4. Run it locally using Ollama or Unsloth AI. This is the shot in the arm that US open source needs! LFG.

Love how @DeepCogito is building its models using self-play - turns out this makes models better both in reasoning and non-thinking modes. Kudos @drishanarora and @DeepCogito team!