Story
October 6, 2026
Beam Is Reflection AI’s High-Stakes Bet to Challenge China’s Open-Model Lead
Reflection AI is pitching Beam as a leaner, open-weight U.S. alternative to Chinese AI models. Its promise of lower inference costs is ambitious, but independent benchmark evidence has yet to catch up.
Reflection AI was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, entering a market where Chinese developers have built a commanding presence in open-weight models. Beam is meant to change that equation—not by matching rivals parameter for parameter, but by making frontier-level AI cheaper to run.1
The startup began assembling the infrastructure for that wager this year. It partnered in May to supply models to the U.S. Department of Energy and Department of War, then signed a $6.3 billion SpaceX compute agreement in June and a further $1 billion deal through Nebius in July. Ahead of Beam’s release, a widely shared post said the model was expected to compete with “the top Chinese open weight ones,” citing Reflection’s heavy spending on Colossus compute.
2
Beam, unveiled this month, is a 501-billion-parameter mixture-of-experts model that activates only 23 billion parameters per token. Reflection says that design gives it competitive coding and agentic performance at a fraction of rivals’ token cost and inference compute. Its announcement called Beam a model that “advances the Western open frontier on coding & agentic tasks,” with full weights due for release this month.3
That pitch lands squarely against China’s scale advantage. Moonshot’s Kimi K3 has 2.8 trillion total parameters and activates 104 billion per token; DeepSeek’s V4-Pro has 1.6 trillion total parameters and activates 49 billion. Reflection says Beam can use up to four times less inference compute, while Laskin argues that “technological progress is driven by values of openness and collaboration.”4
But the contest is not settled by a launch announcement. Reflection has said Beam’s internal tests resemble Z.ai’s GLM-5.2, yet public Artificial Analysis benchmark data was not available at the time of reporting. The company’s efficiency claims now face the harder test: whether developers and enterprises see Beam as more than a patriotic alternative—and as a genuinely competitive workhorse.