Story
October 6, 2026
Beam Bets US Openness Can Catch China’s AI Lead
Reflection AI has launched Beam, a 501-billion-parameter open-weight model designed to narrow China’s lead in open AI by offering competitive coding and agentic performance at a lower compute cost.
Founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, Reflection AI entered a market where China’s DeepSeek, Moonshot and Z.ai had already made open-weight models a strategic force. The New York startup’s answer was to build a Western contender around control, customisation and lower running costs rather than merely bigger model numbers.1
That strategy was backed by a rapid build-out. In May, Reflection partnered to supply models to the US Department of Energy and Department of War; in June, it signed a reported $6.3 billion compute agreement with SpaceX, followed by a $1 billion Nebius deal in July. Ahead of Beam’s debut, one widely shared post said the company had been paying Elon Musk’s Colossus operation $150 million a month for compute since July — a sign of the immense infrastructure behind its open-model push.
2
Beam is the result: a mixture-of-experts model with 501 billion total parameters but only 23 billion activated per token. Reflection says that design gives it an edge over larger Chinese competitors: DeepSeek V4-Pro activates 49 billion parameters per token, while Moonshot’s Kimi K3 activates 104 billion. The company says Beam was trained on 23.8 trillion tokens and can deliver frontier coding and reasoning while using up to four times less inference compute than rival open models.3
The central claim remains one of positioning as much as proof. Reflection says its internal tests put Beam near Z.ai’s GLM-5.2 and calls it a practical “workhorse” for enterprises, developers and public agencies — but independent Artificial Analysis results were not yet available. A launch post described Beam as a “highly efficient agentic open model” with full weights due this month.4
For Laskin, the contest is geopolitical as well as technical: “Open models are Trojan horses for the infrastructure they bring with them.”3 Beam’s Apache 2.0 release is therefore a bid to prevent Chinese open ecosystems from becoming the default global stack — while proving that a US alternative can be cheaper to operate, not simply louder to announce.