AI Revolution: Reflection Unleashes Beam, a Cost-Cutting Open-Weight Model!

Reflection AI, a Brooklyn-based startup established two years ago, has officially unveiled Beam, its inaugural frontier, open-weight AI model. The company asserts that Beam's performance on advanced reasoning benchmarks rivals that of leading Chinese open models like DeepSeek, Qwen, and Z.ai, while significantly reducing costs. This launch is poised to intensify the global competition for developing a prominent Western open-source AI solution.

Details released in a comprehensive blog post confirm that Beam is a text-only mixture-of-experts model. It has been trained using high-compute reinforcement learning to excel in reasoning, coding, and agentic tasks. Reflection AI claims Beam operates at a "fraction of the token cost and inference time compute" compared to its competitors, specifically using "3-4x less inference compute" than rivals for comparable performance.

Beam is a substantial model, featuring 501 billion parameters with 23 billion active parameters. It was pretrained on an extensive dataset of 23.8 trillion tokens and boasts a 1 million token context window. In comparison, Z.ai’s GLM-5.2 has approximately 744 billion total parameters with 40 billion active. While Reflection's performance claims have not been independently verified, the company states Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and surpasses today’s leading Western open models.

Reflection AI positions Beam as a "workhorse model" ideal for enterprises, the public sector, and developers. The company is strategically challenging both closed AI labs such as Anthropic and OpenAI, as well as popular open models from Chinese developers, and Western players like Mistral, Meta, and Cohere. Its most direct U.S. competitor appears to be Inkling, an open model from Mira Murati’s Thinking Machines Lab. Reflection's own benchmarks indicate Beam outperforms Inkling on four coding tests where both models report results, although Inkling is a multimodal model while Beam is exclusively text-only.

Founded in 2024 by two former Google DeepMind researchers, Reflection AI has garnered substantial investment, raising approximately $4.7 billion from prominent backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook. Its most recent funding round valued the company at a $25 billion pre-money valuation. A critical component of its strategy, securing compute resources, has seen Reflection sign deals collectively worth over $7 billion with SpaceX and Nebius to access Nvidia’s GB300 chips through 2029. This compute capacity is essential for training frontier models capable of attracting customers away from established closed models and cheaper Chinese open-weight alternatives.

Reflection AI is targeting enterprises and sovereign nations with Beam and its future models. The company aims to develop "AI factories" – a product enabling institutions to build their own customized, local AI systems by training Reflection’s models on their proprietary data. Nvidia CEO Jensen Huang, a key investor, has long advocated for the "AI factory" concept and a stronger open AI ecosystem, a vision that also benefits Nvidia's GPU business. Hedge funds and trading firms are reportedly among those eager to adopt such systems. Reflection has already commenced testing a sovereign AI factory partnership with Shinsegae Group in South Korea. The company plans to release Beam’s weights and full technical details this month, with distribution facilitated through hyperscalers and neoclouds, alongside integrations across open-source libraries at launch. TechCrunch's requests for additional information were not met by Reflection in time for this report.