On Monday (5), Reflection AI introduced the Beam, its first open-weight model, with 501 billion total parameters and 23 billion active. The company positions the system as an American alternative to the Chinese models that have gained ground in programming, reasoning and tasks performed by AI agents.
Beam uses a Mixture of Experts architecture, which activates only a portion of the network during inference. The strategy reduces the amount of computation required to generate each response and puts efficiency at the center of the competition, instead of relying only on the model's total size.
According to Reflection, Beam delivers performance competitive with GLM 5.2, from the Chinese company Z.ai, and approaches Qwen 3.8 Max in tests focused on programming and agents. Models such as Kimi K3 and DeepSeek V4.1 Flash still appear ahead in different benchmarks released by the company itself.
The results still need to be independently evaluated. Beam is going through the final stages of safety testing and evaluation before the public release of its weights.
Reflection bets on more performance per unit of compute
The main technical bet is on the relationship between performance and computational cost. Reflection claims that Beam achieves results comparable to GLM 5.2 in advanced reasoning tests using three to four times less inference compute.
This comparison is based on estimates of computational operations and the number of tokens generated, not on the actual cost of operating the models in commercial infrastructure. Even so, the indicator directly targets one of the points that helped Chinese labs expand their influence in the open-model market: delivering competitive capability with lower resource consumption.
The development also involved a large-scale reinforcement learning operation. Reflection used 10,500 Nvidia GB300 GPUs for four weeks and generated more than 100 million rollouts, or complete interaction sequences used to train and evaluate the model's behavior. The maximum context reached 256,000 tokens.
Before this stage, Beam was pretrained with 23.8 trillion tokens. The company says its infrastructure was able to support about 110,000 simultaneous rollouts during training.
Reflection plans to release Beam's weights later in October, under the Apache 2.0 license, along with a technical report, model card and tools for execution, evaluation and fine-tuning. The publication will allow researchers and developers to independently test whether the efficiency advantage presented by the company holds up outside its own benchmarks.
