Nvidia-backed Reflection AI challenges Chinese dominance in open-weight models

U.S.-backed startup Reflection AI has introduced an open-weight model named Beam, aimed at challenging China's dominance in freely available AI technology. Beam, designed for coding, reasoning, and agentic tasks, claims to be competitive with GLM-5.2 from Z.ai and approaching Qwen3.8-Max by Alibaba Group. The model employs a "mixture of experts" approach, distributing workloads among specialized networks, which allows it to activate only 23 billion parameters per task out of a total 501 billion.

This significantly reduces compute demands compared to GLM-5.2, which requires 40 billion out of 744 billion parameters. Reflection claims Beam is three to four times more efficient in inference compute than comparable open models, enabling faster and cheaper AI workloads. While official benchmarks are pending, early indicators from Artificial Analysis suggest Beam could be one of the most token-efficient open models in its class.

The launch of Beam follows a push by the U.S. to counter China's rise in open-source AI, with companies like Nvidia, Microsoft, Palantir, Meta Platforms, IBM, OpenAI, and Google signing a letter supporting domestic open-weight AI ecosystems. Reflection AI, founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, has secured over $4 billion in funding and has partnered with the Pentagon, the U.S. Department of Energy, and several U.S. allies.

The model is currently undergoing final testing and will be released to a limited group of users later this month, with full model weights and fine-tuning tools to follow.