Reflection AI, a Brooklyn-based startup founded by former Google DeepMind researchers, has introduced Beam, its first open-weight AI model aimed at competing with prominent Chinese models like DeepSeek, Qwen, and Z.ai. Beam is a 501-billion-parameter text-only mixture-of-experts model, with 23 billion active parameters, pretrained on 23.8 trillion tokens and featuring a 1 million token context window.

The company claims Beam delivers comparable performance to Z.ai’s GLM-5.2 model on advanced reasoning benchmarks while requiring three to four times less inference compute. Reflection positions Beam as a cost-effective option for enterprises, public sector organizations, and developers needing efficient reasoning, coding, and agentic capabilities.

Reflection’s announcement follows reports from Axios and highlights the startup’s ambition to challenge both closed AI labs like Anthropic and OpenAI, as well as open models from Chinese and Western developers such as Mistral, Meta, Cohere, and Mira Murati’s Thinking Machines Lab. Reflection’s own benchmarks indicate Beam outperforms the U.S.-based Inkling model on coding tests, although Inkling supports multimodal inputs while Beam is text-only.

The company has secured substantial funding, raising approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a valuation near $25 billion. Reflection has also locked in over $7 billion in compute agreements with SpaceX and Nebius to access Nvidia’s GB300 chips through 2029, ensuring the resources needed to train and deploy frontier AI models.

Reflection plans to market Beam and future models as part of an "AI factory" concept, enabling enterprises and sovereign nations to build customized AI systems trained on proprietary data. This approach aligns with Nvidia CEO Jensen Huang’s vision of an open AI ecosystem powered by Nvidia GPUs. Reflection is already piloting sovereign AI factory partnerships, including one with South Korea’s Shinsegae Group.

The company intends to release Beam’s model weights and full technical documentation within the month, distributing them through hyperscalers, neocloud providers, and open-source libraries. Reflection has not provided additional comments beyond its blog post and previous reports.