Reflection AI Unveils 501B-Parameter Beam Model with $800M NVIDIA Backing

Reflection AI unveiled Beam on October 7, a 501-billion-parameter open-weight Mixture-of-Experts model positioned as a US answer to Chinese open-weight labs.

Model specifications:

• 501 billion parameters in Mixture-of-Experts architecture

• Trained on 23.8 trillion tokens

• Open weights available

Financial backing:

• $800 million in NVIDIA equity investment

• Compute deal with SpaceXAI reportedly worth $150 million per month through 2029

Reflection AI explicitly positions Beam as competition to Chinese labs including DeepSeek, Alibaba, Moonshot, and Z.ai.

The launch coincided with Mistral's trillion-parameter Large 4 preview and DeepSeek's reported $12 billion funding round — collectively demonstrating that open-weight challengers are closing the capability gap with closed US frontier labs faster than many expected.

Why It Matters: The combination of NVIDIA backing and substantial monthly compute commitment positions Reflection AI to sustain frontier-scale training for years, adding another credible open-weight option for enterprises.