Key facts
- Reflection AI launched Beam, an open-weight AI model designed to compete with Chinese models.
- Beam is a 501-billion-parameter model with 23 billion active parameters and a 1 million token context window.
- The company claims Beam matches leading Chinese open models on reasoning benchmarks at a fraction of the cost.
- Reflection AI has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners.
- The startup has secured over $7 billion in deals with SpaceX and Nebius for compute access through 2029.
- Reflection AI is developing an 'AI factory' product for institutions to build customized local AI systems.
Reflection AI, a two-year-old startup, has officially launched Beam, its first open-weight AI model. The company claims Beam rivals the performance of leading Chinese open models like DeepSeek, Qwen, and Z.ai on advanced reasoning benchmarks, but at a significantly lower compute cost. Beam is described as a text-only mixture-of-experts model trained for reasoning, coding, and agentic tasks.
Beam boasts 501 billion total parameters and 23 billion active parameters, with a context window of 1 million tokens. Reflection AI states that Beam uses 3-4 times less inference compute than its rivals and outperforms current Western open models on certain coding tests, though it is text-only compared to multimodal models like Inkling from Mira Murati's Thinking Machines Lab.
Founded in 2024 by former Google DeepMind researchers, Reflection AI has secured approximately $4.7 billion in funding from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, achieving a $25 billion pre-money valuation in its last round. The company has also committed over $7 billion to deals with SpaceX and Nebius for access to Nvidia's GB300 chips through 2029, aiming to secure the necessary compute power.
Reflection AI is positioning Beam as a "workhorse model" for enterprises, the public sector, and developers, and is developing an "AI factory" product. This product would allow institutions to train Reflection's AI models on their own proprietary data to build customized, local AI systems. Nvidia CEO Jensen Huang supports this vision, which would also benefit Nvidia's GPU sales. Reflection AI is piloting this concept with Shinsegae Group in South Korea and plans to release Beam's weights and technical details this month.