Nvidia-backed AI startup Reflection unveils Beam model

Reflection AI has unveiled Beam, an open-source large language model boasting 501 billion parameters. This follows a recent funding round that valued the startup at $25 billion. Concurrently, it reportedly secured a $6.3 billion agreement with SpaceX to lease Nvidia GB300 NVL72 appliances. These systems, each equipped with 72 graphics cards, were utilized to train the Beam model.

To evaluate its performance, Reflection AI compared Beam with GLM-5.2, an open-source LLM with approximately 250 billion more parameters. The company found that Beam can excel in certain tasks while consuming 25% less hardware. Moreover, Beam's capabilities are on par with Qwen 3.8-Max, a model equipped with over 2 trillion parameters.

This launch is significant as most advanced open-source LLMs have been developed by Chinese firms, including Qwen 3.8-Max and GLM-5.2. Beam stands as the first open-source model from a U.S. startup to demonstrate performance comparable to or surpassing that of frontier models like Anthropic PBC’s Claude Fable 5.1. The creation of Beam began with a small prototype model, which evolved into Beam Base, the foundation upon which Beam is constructed.

Developed using a cluster of 6,144 graphics cards, Beam Base was trained on 23.8 trillion tokens from the public web and commercial sources, including a substantial amount of software code. Custom filters were implemented for each programming language to eliminate low-quality files. The model underwent midtraining, a process that optimized its context window and reasoning abilities.

This stage paved the way for the final phase, which involved training 1.3 billion reinforcement learning sandboxes on 10,000 GB300 graphics cards. These virtual environments were designed to teach Beam new skills, such as code generation, web searching, and AI agent operation. The reinforcement learning phase was completed in just four weeks, aided by software that minimized disruptions due to errors.

Beam is currently accessible through an early access program, with weights, documentation, and fine-tuning tools expected to be released later in the month.