Reflection AI, a Brooklyn-based startup, announced the availability of Beam, its first frontier open-weight large language model. The system is built as a text-only mixture-of-experts architecture and leverages high-compute reinforcement learning to handle reasoning, coding, and agentic tasks. The company says Beam delivers comparable results to leading Chinese models while requiring only a fraction of the token and inference compute.
Beam comprises 501 billion total parameters, of which 23 billion are active during inference, and was pretrained on roughly 23.8 trillion tokens. Its context window stretches to one million tokens, a scale that exceeds many contemporary offerings. For reference, Z.ai’s GLM-5.2 model lists about 744 billion total parameters and 40 billion active parameters, placing Beam in a comparable size class with fewer active weights.
Reflection’s internal benchmarks place Beam on par with Z.ai’s GLM-5.2 on advanced reasoning tests, while consuming three to four times less inference compute. The firm also reports that Beam surpasses the open-source Western leader Inkling on four coding evaluations, despite Inkling’s multimodal capabilities. Independent verification of these results has not yet been published, leaving the broader community to await third-party testing.
Founded in 2024 by two former Google DeepMind researchers, Reflection has attracted roughly $4.7 billion in capital from investors such as Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook. The most recent financing round assigned the company a pre-money valuation of $25 billion. In parallel, the startup secured compute agreements worth more than $7 billion with SpaceX and Nebius, granting access to Nvidia’s GB300 GPUs through 2029.
Reflection positions Beam as a workhorse for enterprises, public-sector agencies and developers, promoting the concept of an “AI factory” that lets organizations train the model on proprietary data. Nvidia’s chief executive Jensen Huang has publicly advocated for such factories, a vision that aligns with Reflection’s plan to distribute Beam through hyperscalers, neoclouds and open-source libraries. Early interest has emerged from hedge funds, trading firms and a pilot partnership with South Korea’s Shinsegae Group.
The company says the full set of model weights and technical documentation will be released later this month, with availability across major cloud providers and integration points for popular open-source toolkits. By offering an open-weight alternative that claims lower compute demands, Reflection aims to attract customers who currently rely on closed offerings from Anthropic, OpenAI or other Western providers, while also providing a competitive option against cheaper Chinese open models.