Reflection unveils Beam, an open model matching GLM-5.2

Reflection's Beam arrives in a market where China's open models, such as Kimi K3 from Moonshot and DeepSeek's V4, have been the most widely used. By claiming benchmark parity with Z.ai's GLM-5.2, the company is positioning Beam as a Western open alternative that can compete on performance. Reflection says it made Beam efficient through high-compute reinforcement learning, delivering competitive results on coding and agentic tasks at a fraction of the token cost and inference time.

The company has moved quickly to secure the infrastructure to support its models. In June it signed a $6.3 billion compute deal with SpaceX for access to Nvidia GB300 chips, and in July it added a $1 billion deal via cloud provider Nebius. It also entered a partnership in May to supply models to the U.S. Department of Energy and the U.S. Department of War, signaling an early focus on public-sector customers.

How much Beam changes the open-model landscape may hinge on whether independent tests confirm Reflection's benchmark and efficiency claims, since the company conducted the benchmarks itself. If they hold up, Beam could become a serious option for enterprises drawn to open models for data control and cost reasons, even as proprietary models from OpenAI, Anthropic, and Google DeepMind continue to lead in performance.