Reflection AI unveils Beam, a US open-weight model aiming to rival Chinese ones while using less compute
Redacción IvokaOct 10, 10:54 AM 3 min
Brooklyn startup Reflection AI unveiled Beam, its first open-weight model. It says it matches Chinese open models on reasoning with far less compute. The weights are due this month.
Reflection AI, a Brooklyn startup backed by Nvidia, unveiled Beam on October 5: its first "open-weight" model, meaning anyone will be able to download and run it themselves once the files are published.
What the company says
Reflection claims Beam matches leading Chinese open models on advanced reasoning benchmarks at much lower compute cost: in its blog, it says it scores comparably to GLM-5.2 while using three to four times less inference compute. It acknowledges that models like Kimi K3 remain ahead on raw capability; its edge, it says, is efficiency.
What it is built for
According to the company, it focuses on coding, reasoning and agentic tasks. It is a text-only model with 501 billion total parameters, of which 23 billion are used per response, and a one-million-token context window.
When you can use it
Beam is still in final safety testing. Reflection says it will release the weights, technical report and developer materials later this month, with distribution through cloud providers and open source libraries.
Why it matters to you
For most creators and small businesses, this is not something you will use tomorrow. Still, it is a release worth keeping on your radar, especially if you follow open models from outside China. The performance figures are the company's own claims.