Beam Bets American Openness Can Close China’s AI Gap
Beam Bets American Openness Can Close China’s AI Gap
Reflection AI was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, entering a market where the biggest Western labs largely kept their most capable models behind APIs. The startup chose the opposite route: open-weight systems that customers can run and modify themselves, betting that businesses will value control over data, security and cost even when proprietary models often lead on raw performance.
That strategy quickly acquired geopolitical weight. Chinese companies including DeepSeek, Moonshot and Z.ai have set the pace among widely available models, while U.S. companies have struggled to field comparable open alternatives. Reflection built capacity for the contest: in May it partnered to supply models to the U.S. Department of Energy and Department of War; in June it signed a $6.3 billion compute agreement with SpaceX; and in July it added a $1 billion deal through Nebius.
Now comes Beam, Reflection’s first open model. The company says its mixture-of-experts architecture contains 501 billion total parameters but activates only 23 billion per token—against 49 billion for DeepSeek V4-Pro and 104 billion for Moonshot’s Kimi K3. The company’s central claim is not that Beam is the biggest model, but that it can deliver frontier coding and agentic performance with up to four times less inference compute.
The launch message was equally direct: Beam is a “highly efficient agentic open model” with 501 billion total parameters and 23 billion active, with full weights promised this month. Reflection says the model is competitive with Z.ai’s GLM-5.2 in its own testing, though independent Artificial Analysis benchmarking was not yet available.
For Laskin, this is about more than a benchmark race. “Technological progress is driven by values of openness and collaboration,” he said, calling those “deeply American values.” The company plans to release Beam under an Apache 2.0 licence and has joined Nvidia’s Nemotron Coalition—an effort to turn that argument into infrastructure, adoption and, ultimately, a Western open-model ecosystem capable of matching China’s.
Write a comment