Reflection Wants to Deliver America’s Most Powerful Open-Weight AI

The arrival of Reflection is cause for self-reflection for the AI frontier labs. The Nvidia-backed US startup said last week that its soon-to-be-released open-weight model Beam rivals the best of the best in the open-source field. It’s big news for domestic AI users who don’t want to pay for costly top-flight closed models from frontier labs but also don’t want to rely on an open-weight ecosystem that hails from China.

Made in the USA

Beam’s launch is a long time coming. Founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis ​Antonoglou, it emerged from stealth last year saying it planned to build powerful open-weight tools intended for business users, with an emphasis on vibe coding and agents. It had massive institutional backing out the gate. Nvidia, a champion of open-source AI, was among the investors in a $2 billion funding round last year that valued Reflection at $8 billion; the company has held talks this year for another fundraising round that would value it at $25 billion, The Wall Street Journal previously reported. It’s also been at the center of the White House’s AI diplomacy plans as it aims to reduce ally reliance on Chinese AI; in March, the US government brokered a deal that will see Reflection aid South Korea in developing its own sovereign AI ecosystem.

Last week, Reflection said Beam compares favorably on advanced reasoning benchmarks to GLM-5.2, the flagship open-weight model by Chinese player Z.AI. That would be better than any US open-weight player, though a tick behind the top Chinese models from the likes of DeepSeek and Moonshot. Crucially, Reflection says Beam requires comparatively less computing power than its Chinese rivals. Those rivals, by the way, are not slowing down

• Bloomberg reported last week that DeepSeek is nearing at least 80 billion yuan ($12 billion) in a new fundraising round, ahead of a planned IPO to come sometime next year. The firm had initially sought just 50 billion yuan for the fundraising round.
• Moonshot, meanwhile, closed a fundraising round that valued it at $50 billion, according to a Bloomberg report. The company is also en route to an early 2027 IPO.

Cheap Shot: Still, the worm may be turning on cheaper open-weight models, which have gained prominence and market share this year in the US. One recent study spanning 6,800 tasks from researchers at Sanford, Carnegie Melon, Berkeley and Microsoft found that cheaper models ended up costing more than more expensive models for one-third of tasks because they took longer to perform and sometimes ended in failure. It’s a lesson we thought no one still had to learn: giving the intern too many responsibilities just creates more work for everyone.