Nvidia-backed US startup Reflection AI introduced Beam, its first open-weight artificial intelligence model, on October 5. The company says the system is designed for coding, reasoning and tool-using agent workloads. Beam uses a sparse mixture-of-experts architecture with 501 billion parameters in total, while about 23 billion are activated for each task. The design is intended to provide high capacity without using the entire network for every request.
According to Reflection’s technical announcement, Beam is a text-only model. The company says its pretraining used 23.8 trillion tokens drawn from web material and licensed datasets. It also reports that the reinforcement-learning stage produced more than 100 million rollouts during four weeks of training on 10,500 Nvidia GB300 processors. These figures come from Reflection and have not been presented as the result of an independent audit.
In a sparse mixture-of-experts system, the model routes each request through only the parts of the network considered relevant to that task. Beam’s use of 23 billion active parameters out of 501 billion total is intended to limit the amount of computation needed during inference. Reflection argues that this allows the model to compete on coding and agentic tasks while reducing cost. Actual operating costs will still depend on hardware, context length and the amount of generated output.
Reflection claims Beam is competitive with Z.ai’s GLM-5.2 and is approaching Alibaba’s Qwen3.8-Max. Reuters, the Financial Times and TechCrunch reported those comparisons as company claims. TechCrunch noted that the performance claims had not yet been independently verified. The announced benchmark results should therefore not be treated as a definitive ranking; they reflect Reflection’s selected tests and evaluation methods.
The launch is part of a wider effort by US developers to compete with Chinese companies in open-weight AI. Open-weight models make trained numerical parameters available so organizations can run the system on their own infrastructure or adapt it to specific uses. The term does not necessarily mean that the training data, source code and every part of the development process are fully open. Licensing and documentation remain important for determining what users can actually do.
Reflection says Beam is still undergoing final red-teaming and evaluations. The company plans to release the model weights, technical report, model card and fine-tuning stack later in October. Early-access registration is available, but the initial announcement does not mean the full package is already generally downloadable. Potential users will need the final documentation to assess licensing terms, safeguards and hardware requirements.
The scale of the reported training effort is also notable. Reflection’s use of a large GB300 cluster and high-compute reinforcement learning shows that open-weight systems can still require substantial energy, hardware and capital. Nvidia’s backing may strengthen the startup’s access to advanced accelerators. That commercial relationship, however, is not a substitute for independent testing of the model’s accuracy, efficiency or safety.
Beam’s significance will become clearer after researchers can inspect the released weights and technical materials. For now, the model stands out for its 501-billion-total, 23-billion-active parameter architecture and its focus on coding, reasoning and agent tasks. If the system performs as claimed, it could provide institutions with another alternative to closed hosted models, but independent evaluations of reliability, cost and security will be decisive after broader access begins.
