In 2024, Reflection AI, a startup with investment from NVIDIA, introduced its inaugural large-scale open-weight model, Beam. This model employs a sparse activation architecture, boasting a staggering 501 billion parameters. However, when tasked with a single assignment, it activates a mere 23 billion parameters, striking an optimal balance between inference speed and operational expenses. Its performance is on par with Zhipu AI's GLM-5.2, and it is gradually closing the gap with Qianwen 3.8-Max in terms of code and agent tasks.

Founded by ex-DeepMind researchers, Reflection AI is primarily focused on developing automation tools tailored for software development. The company aspires to rival Chinese AI models like DeepSeek and Kimi in the realms of code writing and agent tasks. With Chinese open-weight large models gaining a competitive edge in terms of cost-effectiveness, customizability, and code generation prowess, U.S. tech firms are feeling the heat. Significantly, Reflection AI has forged a partnership with SpaceX, securing additional computational resources at its Colossus 2 data center.