Reflection AI releases Beam, a US open-weight AI model

In short: Reflection AI, a US start-up backed by Nvidia, has launched its first open-weight AI model, called Beam, and says it can compete with leading Chinese models.

What happened

Reflection AI said it has released Beam, an “open-weight” AI model. Open-weight means the core settings of the model are shared, so other people and companies can download it, adjust it, and run it on their own computers (like buying a recipe you can tweak at home, instead of only ordering from a restaurant).

The company claims Beam performs about as well as GLM-5.2 from Chinese group Z.ai. Reflection also said Beam is strong at coding and “agentic tasks”, which means AI systems that can take steps to complete a job, such as planning and carrying out a multi-step task, not just answering a question.

Reflection said Beam can do this at a lower “token cost”. Tokens are small chunks of text that AI systems process, and pricing often depends on how many tokens you use. The company has not yet published full technical details to back up its performance claims.

Reflection has worked with the Pentagon and the US Department of Energy, and it has deals to help build AI models for US allies including South Korea. The company has raised more than $4bn since launching in 2024, reaching a reported $25bn valuation before releasing its first model.

Why it matters

Open-weight AI can be cheaper for businesses because they can run it on their own hardware and customize it for their needs, instead of paying ongoing fees to closed systems like ChatGPT or Claude. The US government and some lawmakers also see reliance on Chinese-made models as a national security risk, so a strong US alternative could affect both costs and policy decisions.

Source: Financial Times