Two-year-old startup Reflection AI launches open-weight model Beam, claiming Chinese-level performance at lower cost
Brooklyn-based AI startup Reflection AI has unveiled its first open-weight model, Beam, claiming it matches leading Chinese AI models' performance while using far less compute. The company, backed by Nvidia and other major investors, is targeting enterprises and sovereign nations with the model.
Reflection AI, a Brooklyn-based startup founded two years ago, has publicly unveiled its first frontier, open-weight AI model, named Beam. The company claims it can match the performance of leading Chinese open models, such as DeepSeek, Qwen, and Z.ai, on advanced reasoning benchmarks, while consuming significantly less computing power.
Technical specifications
Beam is a text-only mixture-of-experts model trained using high-compute reinforcement learning to perform well at reasoning, coding, and agentic tasks. The model has 501 billion parameters, with 23 billion active at any time, was pretrained on 23.8 trillion tokens, and supports a context window of up to one million tokens. By comparison, Z.ai's GLM-5.2 has roughly 744 billion total parameters and 40 billion active ones.
Reflection claims Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open models while using three to four times less inference compute. These claims have not been independently verified. In its own coding benchmarks, Beam also outperforms Inkling, the model from Mira Murati's Thinking Machines Lab, though Inkling is multimodal while Beam is text-only.
Funding and strategy
Reflection AI was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, reaching a $25 billion valuation. This summer, the company signed deals worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia's GB300 chips through 2029.
The company is pushing the concept of "AI factories" — a product letting institutions and nations train customized models on their own proprietary data. The idea is championed by Nvidia CEO Jensen Huang. Reflection is already testing such a partnership with South Korea's Shinsegae Group. The company plans to release Beam's weights and full technical details this month.
