The startup Reflection, founded in 2024 by former Google DeepMind researchers, has unveiled Beam, its first open-weights AI model. The text-based model utilizes a mixture-of-experts architecture and was trained using reinforcement learning. It features 501 billion total parameters, with 23 billion active, and supports a 1 million token context window.
Reflection claims that Beam matches the performance of China's GLM-5.2 model in logical reasoning while consuming 3–4 times fewer computational resources during inference. The company also asserts that Beam outperforms the Inkling model from Thinking Machines Lab in coding tasks, though Beam is strictly text-based while Inkling is multimodal.
According to PitchBook, Reflection has raised $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, reaching a pre-money valuation of $25 billion. To secure necessary hardware, the startup signed agreements with SpaceX and Nebius worth over $7 billion for access to Nvidia GB300 chips through 2029.
The company aims to serve corporate and government sectors by enabling the creation of specialized local AI systems. Reflection is promoting a 'sovereign AI' concept, already testing a partnership with South Korea's Shinsegae Group. The model weights and technical specifications are scheduled for release later this month via cloud providers and open-source libraries.
