Reflection's Beam Model Achieves 501 Billion Parameters with High Efficiency

Reflection has announced the launch of Beam, an open-weight model featuring 501 billion parameters, which activates 23 billion parameters per token. This model matches the performance of GLM 5.2 while utilizing three to four times less compute, according to the company. Training for Beam was conducted over four weeks using 10,500 Nvidia GPUs. Reflection, co-founded by Misha Laskin and Ioannis Antonoglou, raised $130 million in seed funding and achieved a valuation of $8 billion following a $2 billion funding round in October 2025. The model's development included a compute-intensive reinforcement learning phase, which led to what Reflection describes as 'emergent capabilities' during training. A technical report, developer documentation, and model weights will be released under the Apache 2.0 license later this month. This follows Reflection's previous launch of Asimov, an agent set to be released in Summer 2025. Competitors in the open-weight model space include Deepseek, Qwen, and Kimi K3.