Reflection AI unveils Beam, a huge free-to-download AI model
Nvidia-backed Reflection AI unveiled Beam on Monday, a 501-billion-parameter AI model whose core files, called weights, it plans to release free later this month. Open models let companies run AI on their own computers instead of renting it. SiliconANGLE said Beam is the first U.S. startup model to rival top Chinese open models.
A U.S. startup is betting that the best AI should be free to download. Reflection AI on Monday introduced Beam, its first "open-weight" model. Weights are the learned settings that make an AI model work. Releasing them means anyone can run Beam on their own hardware.
Beam is large. It has 501 billion parameters in total, but only 23 billion are active at a time, a design that saves computing power. It was trained on 23.8 trillion tokens, or chunks of text and code, and can handle up to 1 million tokens at once, The Next Web reported.
Reflection says Beam is built for coding, reasoning and "agent" tasks, where AI takes actions on its own. It claims Beam competes with GLM 5.2 from China's Z.ai and approaches Alibaba's Qwen 3.8-Max, while using three to four times less computing power. Those scores come from Reflection and have not been independently checked.
SiliconANGLE noted that many of the strongest open models come from Chinese firms. It called Beam the first open model from a U.S. startup to show similar or better results. Free models still trail top paid models, it added.
For now, only a small group on a waitlist can try Beam. Reflection says it will release the weights later this month under an Apache 2.0 license, along with a technical report and tools for fine-tuning. Two government bodies, one in the United States and Britain's AI Safety Institute, are assessing the model, chief executive Misha Laskin told Semafor, according to The Next Web.
Reflection was founded by former Google DeepMind researchers Laskin and Ioannis Antonoglou. It raised money at a $25 billion valuation this year and signed a $6.3 billion deal to rent Nvidia systems from SpaceX, SiliconANGLE reported.
"The only way to own intelligence is, by definition, if it's open," Laskin said.
Open models raise a safety question. Once weights are public, anyone can change them. Reflection says it trained a separate safety model and merged it in, and will publish those results.
What we know: Beam's weights are due later in October.
What we don't: How Beam performs in outside tests.
Where outlets agree SiliconANGLE and The Next Web both report that Beam has 501 billion parameters, is in early access now and will have its weights released later this month.
Where they differ SiliconANGLE says Reflection used 10,000 GPUs for Beam's final training phase while The Next Web says 10,500, and all benchmark claims so far come from the company.