Reflection AI debuts Beam, its open-source 501B model

Reflection AI's launch comes a few months after the startup raised funding at a $25 billion valuation. Around the same time, it reportedly inked a $6.3 billion deal with SpaceX Corp. to rent Nvidia GB300 NVL72 appliances — systems that each contain 72 graphics cards — and it used those systems to train Beam.

The training recipe started with a relatively small prototype model before moving to successively larger, more capable algorithms, eventually producing Beam Base. Reflection AI developed Beam Base on a cluster of 6,144 graphics cards and trained it on 23.8 trillion tokens from the public web and commercial sources, including a significant amount of software code, with custom filters for each programming language to remove low-quality files. A midtraining phase extended the model's context window and enhanced its reasoning, setting up a third, hardware-intensive phase in which 10,000 GB300 graphics cards launched 1.3 billion reinforcement learning sandboxes — virtual environments tuned for generating code, searching the web and running AI agents.

What the outcome hinges on is whether early access users find Beam's day-to-day performance matches the benchmarks. Reflection AI says Beam approaches Qwen 3.8-Max, a model with over 2 trillion parameters, but free LLMs still trail frontier models such as Anthropic PBC's Claude Fable 5.1, so Beam's commercial pull is likely to depend on how the full weights and fine-tuning tools land later this month.