Reflection's Beam matches GLM 5.2 with 3–4× less compute

Reflection is positioning Beam as a Western answer to the open-weight models coming out of China, such as Deepseek and Qwen. The company, founded in 2024 by former Google Deepmind researchers Misha Laskin and Ioannis Antonoglou, raised $2 billion at an $8 billion valuation in October 2025, with Nvidia among the investors, and has signed billion-dollar compute deals with SpaceX and Nebius. Its strategy with Beam is efficiency over raw performance: the model is designed to deliver strong results at low compute cost, targeting businesses that use AI for coding and automated workflows but need to keep operating costs in check.

Beam's capabilities come from a combination of standard large-scale pretraining and a particularly compute-heavy reinforcement learning phase. Reflection says it ran 10,500 Nvidia GB300 GPUs for over four weeks during that RL phase, calling it one of the largest training runs any open lab has done, and that performance kept improving without hitting a ceiling. The company also observed what it calls emergent capabilities, such as Beam getting better at web browsing even though no browsing tasks were part of the training mix. Reflection trained a second model for safety and alignment and merged it with Beam, and plans to publish safety test results and open-source the evaluation methods.

The outcome hinges on whether Beam's efficiency claims hold up for real business workloads and whether the company can close the remaining gap to top open models. Reflection says it is already training a successor that aims to do so, but stronger open models like Kimi K3 still beat Beam on raw performance. For now, Beam's weights are not yet released, and it is still going through final safety testing.