Reflection has launched Beam, a 501-billion-parameter open-weight AI model the company says matches leading Chinese open-weight models on advanced reasoning benchmarks while using a fraction of the inference compute. The release positions a U.S. startup directly against Chinese open models from DeepSeek, Qwen, and Z.ai, as well as against Western open and closed labs.<\/p>\n
Beam is a text-only mixture-of-experts model trained with high-compute reinforcement learning to handle reasoning, coding, and agentic tasks. Reflection says Beam delivers results at a fraction of the token cost and inference time compute of rivals, framing it as a workhorse model for enterprises, the public sector, and developers.<\/p>\n
Beam carries 501 billion total parameters with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and supports a 1 million token context window. By comparison, Z.ai’s GLM-5.2 has roughly 744 billion total parameters with 40 billion active.<\/p>\n
The mixture-of-experts design means only a slice of the model activates for any given query, which is what lets a 501-billion-parameter model run at a fraction of the compute of a dense model of similar scale. Reflection pairs that architecture with reinforcement learning aimed at reasoning and code, the same family of training that lifted open-weight Chinese models to the top of public reasoning leaderboards.<\/p>\n
Reflection says Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and outperforms today’s leading Western open-weight models while using 3x to 4x less inference compute. Those numbers are not independently verified and come from Reflection’s own evaluations.<\/p>\n
On coding benchmarks where both models report results, Reflection’s own benchmarks show Beam outscoring Inkling, the open model from a rival U.S. lab released in July. Inkling is multimodal; Beam is text-only, so the comparison is not direct across every capability.<\/p>\n
Open-weight models release their trained parameters so any organization can run, audit, or fine-tune them. That category has become a strategic battleground. Closed labs such as Anthropic and OpenAI keep model weights private, while Chinese labs have shipped competitive open-weight models at aggressive price points, putting pressure on Western open-weight efforts from Mistral, Meta, and Cohere.<\/p>\n
Beam lands in that gap with a frontier-scale parameter count, a coding-and-reasoning training pipeline, and an inference-cost story aimed at buyers who want open weights without paying frontier reasoning prices. Reflection says it will release Beam’s weights and full technical details this month, distributed through hyperscalers and neoclouds, with integrations across open source libraries at launch.<\/p>\n
Reflection was founded in 2024 and is based in Brooklyn. The company has raised roughly $4.7 billion from backers that include Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a most recent pre-money valuation of about $25 billion, according to PitchBook data.<\/p>\n
The startup has also been stockpiling compute, the key ingredient needed to train frontier models. This summer Reflection signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029.<\/p>\n
Beyond the model itself, Reflection is selling the idea of an AI factory: a packaged system that lets institutions train Reflection’s open-weight models on their own proprietary data inside their own infrastructure. The pitch targets both enterprises and sovereign nations that want to keep data local and customize a model for their own domain.<\/p>\n
Nvidia has publicly championed the AI factory concept, and its investment in Reflection ties that hardware pitch to a software counterpart. Axios has reported that hedge funds and trading firms are among the buyers exploring such systems. Reflection says it has already begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea.<\/p>\n
For organizations evaluating open-weight models, Beam adds a U.S.-backed frontier option with a public inference-cost advantage. The real-world value of that advantage depends on independent benchmark verification, the actual hardware footprint needed to serve the model, and the quality of post-launch support, none of which can be judged from the launch announcement alone.<\/p>\n
The mix of an open-weight release, a compute-heavy training pipeline, and an AI factory product points to a market where model access, training infrastructure, and deployment are sold together rather than as separate layers. Reflection is betting that buyers will pay for the bundle.<\/p>\n
Beam is an open-weight AI model from Reflection with 501 billion total parameters and 23 billion active parameters, designed for reasoning, coding, and agentic tasks. Reflection says it matches leading Chinese open-weight models at a fraction of the inference compute.<\/p>\n
Reflection says Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and uses 3x to 4x less inference compute than rivals, though those claims come from Reflection’s own evaluations and have not been independently verified.<\/p>\n
An AI factory, in the company’s pitch, is a packaged system that lets enterprises and sovereign nations train Reflection’s open-weight models on their own proprietary data inside their own infrastructure. Reflection is already testing a sovereign AI factory partnership with Shinsegae Group in South Korea.<\/p>\n