Brooklyn-based startup Reflection AI has unveiled its first open-weights model, Beam. The model utilizes a mixture-of-experts architecture featuring 501 billion total parameters and 23 billion active parameters. It was trained on 23.8 trillion tokens and supports a context window of 1 million tokens.
The company claims Beam matches the performance of China's GLM-5.2 model while requiring three to four times less compute for inference. Unlike the multimodal Inkling model from Thinking Machines Lab, Beam is strictly text-based. Independent verification of these benchmark results is currently unavailable.
Founded in 2024 by former Google DeepMind researchers, Reflection AI has raised $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, reaching a valuation of $25 billion. The company has also secured contracts worth over $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.
The startup plans to distribute Beam via hyperscalers and cloud providers, with weights and documentation scheduled for release in October. The model is positioned for enterprise, government, and developer use. A pilot program for fine-tuning the model on proprietary data is currently underway with South Korea's Shinsegae Group.
