Reflection AI has unveiled Beam, its first frontier open-weight artificial intelligence model, positioning it as a cost-effective alternative to leading Chinese models. The Brooklyn-based startup, founded in 2024 by former Google DeepMind researchers, announced that Beam matches the performance of prominent Chinese open models on advanced reasoning benchmarks while requiring significantly less computational power for inference.

According to the company's blog post, Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters. It was trained on 23.8 trillion tokens and supports a 1 million token context window. The company states that Beam achieves comparable results to Z.ai's GLM-5.2 (744 billion parameters, 40 billion active) on reasoning benchmarks but uses 3-4 times less inference compute.

Reflection emphasizes Beam's efficiency for enterprise and government use cases, describing it as a "workhorse model" for tasks including reasoning, coding, and agentic workflows. The company is directly competing with Chinese developers like Z.ai and Qwen, as well as Western open-weight models from Mistral, Meta, and Cohere. Internal benchmarks indicate Beam outperforms Inkling, a multimodal model from Thinking Machines Lab, on specific coding tests where both models report results.

To support its ambitions, Reflection has secured over $7 billion in compute agreements with SpaceX and Nebius for access to Nvidia's GB300 chips through 2029. The startup is promoting its "AI factory" concept, enabling institutions to build customized AI systems by training Beam on proprietary data. Nvidia CEO Jensen Huang has publicly endorsed this vision, which aligns with Nvidia's interest in expanding its GPU ecosystem.

Reflection plans to release Beam's weights and technical documentation later this month, with availability through hyperscalers, neoclouds, and open-source libraries. The company has not yet responded to requests for additional details.

The startup has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, reaching a $25 billion pre-money valuation. Hedge funds and trading firms have shown particular interest in adopting Reflection's sovereign AI factory approach, as demonstrated by ongoing discussions with South Korea's Shinsegae Group.