Reflection AI

Market position

Reflection is described as an open-source-oriented frontier lab intended as an American alternative to Chinese open models such as DeepSeek and Qwen, and as an open counterpart to closed labs including OpenAI and Anthropic. Its valuation moved from about $545 million in March 2025 to $8 billion in October 2025 and $25 billion as of June 2026. Its October 2025 raise drew public support from figures including White House AI and Crypto Czar David Sacks and Hugging Face CEO Clem Delangue.

The company commits to publishing open model weights, research papers and open-source customization tooling, arguing that open systems innovate faster and are safer because external researchers can inspect weights and probe risks. It targets enterprises and sovereign customers that prefer models they can own, self-host, cost-control and customize, and that may avoid Chinese models for legal reasons.

Technology

The company's work spans large language model training, reinforcement learning, agentic AI and software engineering automation, with emphasis on reinforcement learning at scale and on both pre-training and post-training systems. In October 2025 it said it had built a large-scale LLM and reinforcement learning platform capable of training Mixture-of-Experts models at frontier scale, first applied to autonomous coding and then extended to general agentic reasoning. Its Asimov agent performs code comprehension over source code and workplace communication and documentation sources.

Go-to-market

Reflection combines open releases of model weights and research with direct enterprise and government agreements and infrastructure partnerships. Disclosed collaborations include Dell for on-premises AI factory deployments, the U.S. Department of Energy for the National Laboratories and Genesis Mission, and Shinsegae for a sovereign South Korean cloud. Large compute supply agreements with Nvidia-based providers underpin its model training roadmap.

Large enterprises that want ownership, on-premises deployment, cost control and customization of models; national governments pursuing sovereign AI; national laboratories and public research programs; and the broader research and developer community using open weights.

Geography

Headquartered in Brooklyn, New York, in the United States, with announced activity in South Korea via a 250-megawatt data center project with Shinsegae and U.S. federal engagements through the Department of Energy and National Laboratories.

History

Founded in 2024, the company emerged from stealth in March 2025 with $130 million in total financing at a roughly $545 million valuation. In July 2025 it launched Asimov, a code-comprehension agent. In October 2025 it announced a $2 billion round at an $8 billion valuation, disclosed the recruitment of researchers from DeepMind and OpenAI, and described a large-scale LLM and reinforcement learning platform capable of training Mixture-of-Experts models at frontier scale. During 2026 it signed infrastructure and government agreements, including a memorandum of understanding with Shinsegae for a 250-megawatt data center in South Korea (March 2026), a role as AI model provider to the U.S. National Laboratories under the Department of Energy's Genesis Mission (May 2026), a compute agreement tied to the SpaceX Colossus data center (June 2026), and an agreement with Nebius for more than $1 billion of computing power (July 2026). It was reported in March 2026 to be in talks to raise $2.5 billion at a $25 billion pre-money valuation, and was valued at $25 billion as of June 2026.

Risks & controversies

The company carried a $25 billion valuation as of mid-2026 despite reporting that it had yet to generate substantial revenue and had not released its first foundation model as of October 2025, and it has committed to multi-year compute obligations including $150 million per month through 2029 and over $1 billion with Nebius. Its openness is limited to model weights, with datasets and full training pipelines kept proprietary, and observers noted that it still needs to demonstrate a high velocity of open model and dataset releases.