While the headlines fixate on closed models like GPT or Claude, the real fight for AI's future is happening in a different corner of the ring — open-weight models anyone can download and run. For nearly two years, China has been winning that fight: DeepSeek, Alibaba's Qwen and Moonshot's Kimi K3 have consistently outpaced Western rivals in open-source quality. On October 5, Brooklyn-based Reflection AI tried to flip the script with a new model called Beam.
Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers. In just over a year the company raised roughly $4.7 billion and was valued at $25 billion as of April. Its backers include Nvidia, Sequoia and Lightspeed — money that isn't accidental, since Nvidia has an obvious stake in making sure strong open models don't come exclusively out of Beijing.
Beam is a mixture-of-experts model with 501 billion total parameters, only 23 billion of which activate at once, trained on 23.8 trillion tokens with a one-million-token context window. By Reflection's own benchmarks, Beam matches Z.ai's GLM-5.2 — one of the strongest open models on the market today — while using three to four times less compute. The company also claims Beam beats Thinking Machines Lab's multimodal Inkling model on several coding benchmarks.
One caveat matters here: none of these numbers have been independently verified yet. Reflection plans to release the weights and technical documentation publicly later in October, which will let outside developers test the claims on their own workloads.
Why this matters beyond the specs: enterprises and governments increasingly prefer open models because they can run them on their own infrastructure instead of shipping sensitive data to a third party. But as long as Chinese labs make the strongest open models, they set the bar everyone else — including the banks and agencies wary of sending data to China — has to follow. Beam is a bet that the US can field a serious open-weight contender, not just win the closed-model race.
Reflection already has compute lined up for what comes next: a SpaceX deal for Nvidia GB300 chip access worth more than $6 billion, plus a roughly $1 billion contract with cloud provider Nebius. The company says it's already training a larger model for its next release.
