Reflection AI Launches Beam: A High-Performance Open-Weight Model at Lower Compute Costs
Reflection AI has launched its first open-weight model called Beam, claiming it matches the performance of top Chinese models like Z.ai’s GLM-5.2 but with significantly lower compute costs and faster inference times. The text-only model features a massive scale—501 billion parameters—with 23 billion active ones—and was trained on over 23 trillion tokens. Reflection positions Beam as an alternative to both closed AI labs such as Anthropic and OpenAI, as well as Western open models like Mistral or Meta’s offerings. The startup has raised $4.7 billion in funding from major investors including Nvidia and Sequoia Capital, with a pre-money valuation of $25 billion. To support its model development, Reflection secured over $7 billion worth of compute deals for access to NVIDIA's GB300 chips through 2029. The company is also promoting the concept of 'AI factories'—customized AI systems built on proprietary data—which aligns with Nvidia CEO Jensen Huang’s vision and could attract enterprises and sovereign nations. Beam will be released this month, including full technical details and weights, distributed via hyperscalers and open-source libraries. Reflection aims to compete directly in the growing market for efficient frontier models while offering a Western alternative to Chinese AI advancements.
Tickers: $META
Topics: Artificial Intelligence Model Performance Open-Weight Models Startup Launches