Beam Is Reflection AI’s Bid to Break China’s Open-Model Lead
Beam Is Reflection AI’s Bid to Break China’s Open-Model Lead
Reflection AI, founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, entered a race in which Chinese developers have set the pace for popular open models, including DeepSeek and Moonshot AI’s Kimi. The New York company argues that the contest is no longer simply about who builds the biggest model, but who can make powerful systems practical to run.
That argument has been backed by an aggressive build-out. In May, Reflection partnered to supply models to the U.S. Department of Energy and Department of War; in June, it signed a reported $6.3 billion compute agreement with SpaceX, followed by a further $1 billion arrangement through Nebius in July. Before Beam’s debut, an X post amplified reports that Reflection was preparing an open-weight model capable of competing with China’s leaders.
The launch turns that ambition into a concrete claim. Beam is a 501-billion-parameter mixture-of-experts model, but activates only 23 billion parameters for each token. Reflection says that compares with 49 billion for DeepSeek V4-Pro and 104 billion for Kimi K3, allowing Beam to deliver frontier coding and reasoning performance with up to four times less inference compute.
The company’s case rests on efficiency rather than raw scale. Kimi K3 has 2.8 trillion total parameters and DeepSeek V4-Pro has 1.6 trillion, underscoring the size advantage of the Chinese systems Reflection is targeting. Yet Reflection says high-compute reinforcement learning made Beam “extremely efficient at reasoning,” with competitive coding and agentic performance “at a fraction of the token cost and inference time compute.”
Beam’s full weights are due this month, under an Apache 2.0 license. Reflection is positioning it as a deployable workhorse for enterprises, developers and government users seeking control over proprietary data and local infrastructure—not merely another benchmark contender. Its announcement framed the model as having “501B total parameters and 23B active” and advancing the Western open frontier on coding and agentic tasks.
For Laskin, the strategic stakes are explicit: “Open models are Trojan horses for the infrastructure they bring with them.” Beam is therefore both a model launch and a bid to ensure that the infrastructure behind the next open-AI wave is not overwhelmingly Chinese.
Write a comment