Nvidia-backed Reflection Unveils Beam: Its First Open-Weight Model
Reflection AI, the Nvidia-backed startup founded by former Google DeepMind researchers, has unveiled Beam, its first open-weight model. Beam is built for coding, reasoning, and agentic work. The company plans to release its weights later this month, as mentioned in a blog post on Monday.
What Reflection Says Beam Can Do
Beam is a mixture-of-experts model with 501 billion parameters in total, but only 23 billion are active at any given time. Pretrained on 23.8 trillion tokens, its context window reaches one million tokens. Reflection claims Beam matches China’s GLM-5.2 on reasoning tasks using three to four times less compute. Although Moonshot AI’s Kimi K3 remains ahead in raw capability, Beam excels in efficiency.
Training and Safety
Reflection trained Beam in under four weeks on 6,144 Nvidia (find out more) GB300 GPUs, followed by a four-week reinforcement learning run on 10,500 GPUs. This process yielded over 100 million attempts at various tasks. The company believes this is one of the largest such training runs conducted by any open lab.
During training, Beam independently learned to browse the web and utilize text-recognition tools to read documents, even though these tasks were not explicitly included.
Reflection has signed several compute deals, including a $6.3 billion deal with SpaceX and a $1 billion deal with Nebius. Two government bodies are currently assessing the model with Reflection, namely the US Center for Advancing Innovation and Standards for Super Intelligence and the UK’s AI Safety Institute.
Future Plans
Beam will be released under an Apache 2.0 license this month, along with a technical report, a model card, and tools for running and fine-tuning.