Reflection AI unveils Beam, an open 501-billion-parameter model said to match China's GLM-5.2 with 3-4 times less compute

Nvidia-backed US startup Reflection AI presented Beam on Monday, October 6 - a mixture-of-experts model with 501 billion parameters, 23 billion of them active at a time. According to the company, it matches China's GLM-5.2 in reasoning with 3-4 times less compute and comes close to Qwen 3.8-Max on some coding tasks, although in its own tests it still loses to some Chinese models. The model's weights, along with tools for running, evaluating and fine-tuning it, are due to be released in October; in the future, the company wants to sell an "AI factory" - software in which businesses build their own models on their own data instead of paying for ChatGPT or Claude. Reflection rents its compute from SpaceX: from July 1, 2026, through the end of 2029, it pays $150 million a month for Nvidia GB300 cards at the Colossus 2 data center - about $6.3 billion in total, although after the first three months either side can terminate the contract with 90 days' notice. Nvidia has invested $800 million in the startup.