Beam: Reflection AI Takes Aim at China’s Lead in Open A.I. Models

Reflection AI is trying to do something unusually ambitious for an American start-up: beat China’s open A.I. labs at their own game. The company, founded by two former DeepMind researchers and backed by Nvidia, has unveiled Beam, its first open-weight model. On paper, it is huge. Beam has 501 billion parameters, yet only 23 billion are active for each token, which is how Reflection says it can run much more efficiently than its size suggests.

Beam is text-only, but it can handle a context window of up to one million tokens, and users can dial in how much effort it spends on an answer. Reflection says the model is still in final red-teaming and evaluation. A selected group can already get early access through a waitlist, and the weights are supposed to arrive later this month under Apache 2.0, along with a technical report, a model card and tools for running and fine-tuning it.

The training story is the other part of the pitch. Reflection says it pre-trained Beam on 23.8 trillion tokens drawn from the web, public sources and licensed datasets. The base run used 6,144 Nvidia GB300 GPUs for less than four weeks. Then came a reinforcement learning run of the same length on 10,500 GB300 chips, with more than 100 million rollouts across roughly one million training environments. The company calls that one of the largest open-lab reinforcement learning runs so far, and says the model was still getting better when training stopped.

But for now, all of the benchmarks come from Reflection itself, so the bravado is doing a lot of work. The company says Beam matches Z.ai’s GLM-5.2 on demanding reasoning tasks while using three to four times less compute at inference, and that its edge grows against models above two trillion parameters such as Alibaba’s Qwen 3.8 Max. In its own comparisons, Beam lands near or behind stronger Chinese models on several tests, and it is still not listed in the Artificial Analysis Intelligence Index. Reflection is asking people to admire the engine before they can open the hood. That’s bold, or reckless, depending on how much patience you have for startup theater.

The funding side is just as eye-catching. Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, who worked on Gemini and AlphaGo at Google DeepMind. It raised $2 billion last year at an $8 billion valuation, and The Wall Street Journal says it is now valued at around $25 billion. The company has also been called the “DeepSeek of the West,” which is a flattering label if you like pressure. Beam is only the first model in the series, and its successor is already training.