Reflection Unveils Beam Open-Weight AI Model to Challenge Rivals at Lower Compute Costs

Brooklyn-based startup Reflection AI has officially unveiled Beam, its first frontier, open-weight artificial intelligence model. Founded in 2024 by former Google DeepMind researchers, the company claims Beam matches the performance of leading Chinese open-weight models on complex reasoning benchmarks while slashing inference compute requirements. The launch signals an escalating race by Western developers to offer direct, cost-efficient open alternatives to dominant systems from DeepSeek, Qwen, and Z.ai.\n\nBeam is built as a text-only mixture-of-experts model boasting 501 billion total parameters, with 23 billion active parameters per token, and features a 1 million token context window. Pretrained on 23.8 trillion tokens, the model was refined using high-compute reinforcement learning for reasoning, coding, and autonomous agent tasks. According to company disclosures, Beam operates at “a fraction of the token cost and inference time compute” of its rivals, performing on par with Z.ai’s GLM-5.2 while using three to four times less inference compute.\n\nBacked by substantial investor capital, Reflection has raised approximately $4.7 billion to date from prominent institutions, including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. PitchBook data places the startup's latest pre-money valuation at $25 billion. To guarantee long-term infrastructure, Reflection recently secured compute agreements exceeding $7 billion combined with SpaceX and Nebius, locking in critical access to Nvidia GB300 chips through 2029 to supply enterprise-scale workloads.\n\nPositioning Beam as a "workhorse model," Reflection is targeting sovereign governments and enterprise clients looking to establish localized "AI factories." The firm is already piloting sovereign infrastructure initiatives with South Korea’s Shinsegae Group, alongside strong demand from hedge funds and trading firms. Reflection plans to release Beam’s weights, comprehensive technical documentation, and integrations across hyperscalers and open-source ecosystems later this month.