Deep Cogito
Category: Foundation Models / LLMs
Post-training research lab building self-improving, open frontier AI models using reinforcement learning and Iterated Distillation and Amplification (IDA). Deep Cogito was founded in 2024. The company is led by Drishan Arora. Based in San Francisco, United States. Team size: 11-50. Total funding raised: $56.4M. Latest round: Series A. Key investors include TQ Ventures, Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, Zscaler.
- Founded
- 2024
- Headquarters
- San Francisco, United States
- Team size
- 11-50
- Total funding
- $56.4M
Value proposition
Deep Cogito builds the 'post-training engine' that takes pre-trained base models (including open models like DeepSeek) and applies iterative self-improvement techniques to unlock reasoning and capability gains beyond pre-training alone, offering both open-source frontier models and enterprise customer-specific specialization.
Products and solutions
Cogito v1 and v2/v2.1 model families (3B to 671B MoE, open-licensed, hybrid reasoning models), enterprise post-training/customization services.
Unique value
Uses Iterated Distillation and Amplification (IDA) to run self-play reinforcement learning on open base models (e.g. forked from DeepSeek-V3), producing state-of-the-art open models while offering enterprise-specific post-training as a service.
Target customer
Enterprises needing customized frontier-grade AI models, developers/researchers using open-source LLMs, cybersecurity and cloud platforms (e.g. Zscaler).
Industries served
Foundation Models / LLMs, Enterprise AI customization, Cybersecurity (via Zscaler partnership)
Technology advantage
Proprietary IDA (Iterated Distillation and Amplification) self-improvement method; deep post-training expertise from founders' prior work leading Google Gemini's AI Search post-training; models proven from 3B to 671B+ parameter scale.
How they differentiate
Focuses exclusively on post-training/self-improvement rather than pre-training from scratch, achieving frontier performance on top of open base models like DeepSeek-V3 at lower cost, with transparent benchmarking and enterprise customization offering.
Main competitors
DeepSeek, Mistral AI, Together AI (as model ecosystem peer), other open-model post-training labs.
Key partnerships
Zscaler (customer turned strategic investor, part of AI-Guardian alliance alongside AWS, CoreWeave, Databricks, Google Cloud, OpenAI), model distribution via Hugging Face, Together AI, Fireworks AI, Baseten, RunPod, Ollama, OpenRouter
Notable customers
Zscaler (enterprise customer and strategic investor)
Major milestones
Founded June 2024, $13M seed Aug 2025 led by Benchmark, stealth launch of Cogito v1 (April 2025) on Llama 3.2 base outperforming larger models, Cogito v2/v2.1 release (671B MoE forked from DeepSeek-V3 base), $43M Series A Aug 2026 led by TQ Ventures, total funding >$56M.
Growth metrics
Emerged from stealth April 2025 with Cogito v1 models; released Cogito v2/v2.1 (671B MoE) topping open-model benchmarks; expanding enterprise post-training business post-Series A.
Market positioning
Positioned as a U.S.-based open-source alternative in the post-training/RL specialization niche, competing with both Chinese open model labs (DeepSeek) and closed frontier labs on reasoning benchmarks.
Geographic focus
Global, with emphasis on U.S. enterprise market and open-source AI community worldwide.
About Drishan Arora
Ex-Google, led Gemini's post-training for AI Search (AI Mode/AI Overviews); senior software engineer background; Columbia University alum.
Latest news about Deep Cogito
More Foundation Models / LLMs companies
Official website: https://www.deepcogito.com