Reflection AI has launched Beam, an open-weight AI model with 501 billion parameters, aimed at competing with Chinese AI products.
Reflection AI, an American startup, has launched Beam, its first open-weight AI model featuring an impressive 501 billion parameters, with 23 billion activated per query.
The debut of Beam signifies Reflection AI's expansion beyond its internal projects, positioning it as a competitor to notable Chinese models like Z.ai's GLM-5.2 and Alibaba's Qwen3.8-Max. The startup claims that Beam can achieve similar performance levels to GLM-5.2 while maintaining significantly reduced inference costs.
Utilizing a Mixture of Experts (MoE) architecture, Beam's design means that while it houses 501 billion parameters, only those 23 billion are active during a single request, allowing for optimized computational efficiency. Reflection asserts that Beam's performance in advanced reasoning tasks matches that of GLM-5.2, yet uses 3 to 4 times less computational power for inference. However, the company concedes that the Chinese Kimi K3 model currently excels beyond Beam in overall performance.
Massive Training Infrastructure
Beam's development involved extensive resources, relying on a robust Nvidia infrastructure. The training process spanned four weeks, utilizing 10,500 Nvidia GB300 GPUs and conducting over 100 million training runs. Furthermore, around 1.3 billion virtual environments were utilized during training, which reached a maximum context length of 256,000 tokens and supported up to 110,000 parallel executions simultaneously.
The model was pre-trained on a cluster of 6,144 Nvidia GB300 NVL72 GPUs using 23.8 trillion tokens sourced from the web, public data, and licensed datasets, accumulating substantial training data in a short time.
Beam's Accessibility and Future Outlook
Currently, Beam is available through a waitlist as it undergoes final security evaluations. Early access is granted to a select group of pre-registered users. Reflection plans to release the model weights, a technical report, and developer tools under the Apache 2.0 license in the upcoming weeks, enabling users to deploy, tweak, and integrate the model into their applications.
In context, major American companies, including Nvidia, Meta, and Microsoft, previously advocated against restrictions on open-weight models, emphasizing the importance of innovation in AI technology.