Deep Cogito Inc.

Deep Cogito Inc. is a San Francisco frontier AI research company founded and incorporated in June 2024 by former Google colleagues Drishan Arora and Dhruv Malhotra. Arora, CEO, studied electrical engineering at IIT Delhi and computer science/machine learning at Columbia, then became a senior Google software engineer and tech lead on question answering for Search/Assistant before leading language-model work for generative search. Malhotra, product leader/co-founder, worked across Microsoft Office, Google Search and Google DeepMind, where he translated NLP/ML research and generative-search technology into products. They formed Deep Cogito around a contrarian model-development thesis: many reasoning systems improve by spending more inference tokens searching, but a stronger route to superintelligence is to distill discoveries from search back into the model's parameters so the model develops better 'intuition' and starts the next search closer to the answer. The company calls its framework Iterated Distillation and Amplification (IDA) and iterative policy improvement. It aims to create training signals that can scale beyond the intelligence of a human overseer, and publicly commits to releasing its models as open weights. Deep Cogito emerged from stealth on April 8, 2025 with Cogito v1 Preview: hybrid text models at 3B, 8B, 14B, 32B and 70B parameters derived from Meta Llama and Alibaba Qwen bases. Each checkpoint could answer directly or expose a self-reflective reasoning phase, supported 128k context and more than 30 languages, and targeted coding, STEM, instruction-following, function calling, tools and agents. The company said the small team built the family in roughly 75 days and outperformed size-matched open models on selected benchmarks; those are developer/company benchmark claims rather than independently audited general superiority. On July 31, 2025 it released Cogito v2 Preview at 70B dense, 109B mixture-of-experts, 405B dense and 671B MoE. The largest used an open-licensed DeepSeek base; the others used Llama-family bases. Deep Cogito said all eight v1/v2 models, data generation and more than a thousand training experiments cost under $3.5m combined, and that the 671B reasoning chain was 60% shorter than DeepSeek R1 0528 while performing comparably on its chosen tests. It also acknowledged that its models had not yet received the post-training compute of major frontier labs, that public benchmarks do not capture real-world behavior and that closed models such as o3 and Claude 4 Opus remained stronger. Cogito v2.1 671B, released November 19, 2025, refreshed the DeepSeek-derived model with process supervision, reinforcement/self-play and improved instruction following, coding, long queries, multi-turn interaction and creativity. The company called it the strongest U.S. open-weight model and made weights available on Hugging Face plus API access through OpenRouter, Fireworks AI, Together AI, Ollama Cloud, Baseten and RunPod. Deep Cogito also runs a free chat interface and states that it does not retain chats beyond operational/abuse needs or use inputs for training. The product is therefore both research artifact and ecosystem strategy: developers can download/commercially use weights subject to the underlying model licenses, run them locally, or pay inference partners. Deep Cogito has not disclosed its own paid API, enterprise contracts, revenue, ARR, margins or customers. Its likely near-term economic model is venture-funded research with possible future enterprise inference, hosted access, custom models, licensing/support or partnerships; today, third-party inference providers capture the visible usage fees. For example OpenRouter lists Cogito v2.1 at $1.25 per million input tokens and $1.25 per million output tokens, a partner price rather than company list pricing. Funding has two confirmed layers. A $13m seed announced in August 2025 was led by Benchmark/Eric Vishria with South Park Commons/Aditya Agarwal, both thanked by the company in its v2 research release. Some databases list Innovation Endeavors and angels Amjad Masad, Gokul Rajaram and Spencer Kimball, but no primary or sufficiently direct round evidence was found, so they are unresolved rather than asserted. A 2026 Regulation D equity offering began March 24: the April 8 filing reported $28,524,961 sold and an April 24 amendment raised the sold amount by $11,474,995 to $39,999,956, essentially the full $40m offering, from seven investors. Directors named in the filing are founders Arora/Malhotra, Benchmark partner Vishria and TQ Ventures co-founder Schuster Tanger, corroborating Benchmark and TQ relationships but not round check sizes or lead status. Exact disclosed funding is therefore $52,999,956, conventionally $53m, not $13m. No valuation was filed. As of August 2026 Deep Cogito remains an active private company with a very small research/engineering footprint: its Hugging Face organization lists 14 members and a secondary company profile reported about 11 employees; neither is an audited headcount. The privacy policy says the company operates in the United States and India. Distribution includes Hugging Face, Ollama, Unsloth, Together AI, Baseten, RunPod, Fireworks, OpenRouter and local frameworks, while support/acknowledgements also name Nebius, LM Studio, Anthromind, Llama and DeepSeek teams. Competitive peers span open-model labs Meta, DeepSeek, Qwen/Alibaba, Mistral, AI2, NVIDIA and Nous Research and closed labs OpenAI, Anthropic, Google DeepMind and xAI, as well as reasoning/post-training companies. Differentiation is a lean lab focused on distilling reasoning into a better policy rather than only extending chain length, a single model that toggles direct/reasoning behavior, broad open-weight distribution and claimed training efficiency. The trade-off is limited proprietary control: open weights accelerate adoption and scrutiny but make direct monetization and defensibility harder, while Deep Cogito depends on third-party base-model licenses, cloud/GPU providers and hosting partners. Technical and diligence risks are substantial. Benchmark results may be overfit, contaminated or not representative; shorter chains do not necessarily mean safer or more accurate reasoning; chain-of-thought visibility can mislead users into trusting plausible rationales. IDA/superintelligence claims are research hypotheses, not proven scalable guarantees. Model post-training data provenance, synthetic-teacher dependencies, labor/data rights and base-model-license compatibility are not fully disclosed. Forking DeepSeek/Llama/Qwen imports upstream architecture, training-data, export-control, supply-chain and license risks. Open-weight frontier capability can enable malware, weapons knowledge, fraud, impersonation, disinformation and removal of safety controls; the website terms prohibit harmful use but downloaded weights cannot be centrally revoked. Models can hallucinate, produce bias, leak memorized data and provide unsafe medical/legal/financial answers. The company's free chat routes inputs to third-party infrastructure and collects telemetry/approximate location, though policy promises data minimization and no training on inputs. Its terms prohibit extracting weights from the hosted service or using it to train a competitor while separately releasing weights, making the hosted-service license distinct from model artifact licenses. No material company-specific litigation, regulatory action, security breach or safety incident was found in the manual review. The company name is easily confused with unrelated Cogito enterprise call-center AI, Cogito Tech, academic Cogito/VL-Cogito/Audio-Cogito projects and other software; those facts must not be attached. Key unresolved matters are valuation/cap table, full 2026 investor syndicate, real revenue and customers, independent safety/evaluation data, training data provenance, current headcount, compute commitments and the next release after v2.1.

Founding story

Arora and Malhotra left Google/DeepMind in June 2024 to pursue open superintelligence. Drawing on generative-search and language-model product experience, they built a lean lab around the belief that repeated search should be distilled into better model intuition rather than merely paid for again at every inference.

Business model

Venture-funded frontier model research lab distributing commercially usable open-weight language models through local downloads and third-party inference platforms, with future hosted/enterprise monetization not yet disclosed.

No company revenue model verified. Current models/chat are free or open-weight; inference partners charge usage fees. Potential future hosted API, enterprise support/customization/licensing or co-development is plausible but not announced as material revenue.

Traction

Released five v1 models and four v2 models plus v2.1 within seven months; Hugging Face organization has 14 members and 566 followers as observed Aug 2026; broad distribution across at least seven inference/local platforms. Company claims all eight v1/v2 models cost under $3.5m to train and v2 uses 60% shorter reasoning chains than DeepSeek R1 on its tests. No revenue, paid customers, API volume, downloads, retention or production adoption disclosed.

Latest developments

A March-April 2026 SEC offering sold $39,999,956 of equity to seven investors, lifting exact disclosed funding to $52,999,956. May 2026 terms/privacy formalized free chat data practices. The official research index still lists Cogito v2.1 671B (Nov 2025) as the latest model through Aug 17, 2026; current site continues recruiting for LLM and AI infrastructure roles.