Reflection AI Unveils Beam, a 501B Open-Weight Model Built for Efficient Reasoning

Reflection AI has officially introduced Beam, its first open-weight frontier model, as the startup tries to position itself as a U.S. answer to the strongest open models coming out of China.

The launch follows weekend reporting from Axios that Reflection was nearing the release of a new open model. In a detailed announcement on Monday, the company said Beam is a text-only mixture-of-experts model built for reasoning, coding and agentic workloads, with an emphasis on delivering strong performance at lower inference cost.

Beam is large, sparse and built for efficiency

Reflection says Beam has 501 billion total parameters, with 23 billion active parameters, making it sparse rather than activating its entire network for every request. The startup also says Beam was pre-trained on 23.8 trillion tokens and supports a 1 million-token context window, putting it in the same frontier category as some of the most ambitious open models now competing for enterprise and developer adoption.

The company’s central claim is not just scale, but efficiency. Reflection says Beam performs at a level comparable to leading Chinese open models on advanced reasoning benchmarks while requiring significantly less inference compute. Those claims have not yet been independently verified, but they clearly target enterprises and governments that want frontier-level capability without the full cost profile of rival models.

Reflection is aiming at both Chinese and Western rivals

The launch comes as Western AI companies face increasing pressure from Chinese open-model developers such as DeepSeek, Qwen and Z.ai. Reflection is also competing more directly with Western open-model efforts from companies such as Meta, Mistral and Cohere, while trying to distinguish itself from closed-model leaders like OpenAI and Anthropic.

The broader strategic backdrop is one Reflection has discussed before. In a CNBC interview earlier this year, the company’s chief executive argued that the best open models were increasingly coming from China, underscoring why U.S. companies needed to respond with stronger open alternatives of their own.

The bigger pitch is sovereign and enterprise AI infrastructure

Beam is also part of a larger business strategy. Reflection wants to supply enterprises and sovereign institutions with the foundation for custom AI systems trained on proprietary or national data. Rather than offering only a model for developers to download, the company is pitching what it describes as AI factories: local or dedicated AI systems that organisations can adapt to their own needs.

Reflection has already started testing that approach internationally. In a previously announced partnership with Shinsegae Group in South Korea, the company said it would help build a Korean sovereign AI factory, offering an example of how it hopes to deploy its models and infrastructure at national or institutional scale.

Beam could sharpen the open-model race

Reflection says Beam’s weights and fuller technical details will be released later this month, with distribution through hyperscalers, neocloud providers and open-source software libraries. If those releases arrive as promised, Beam could quickly become an important test of whether a well-funded U.S. startup can narrow the performance and efficiency gap with the open models that have recently helped China gain momentum in the global AI race.

For now, Beam’s significance lies as much in what it represents as in its raw specifications. Reflection makes a clear argument that the next phase of the model race will not be decided only by benchmark scores, but by who can deliver frontier performance, lower compute demands, and a deployment model that appeals to businesses and governments alike.