Reflection AI launches Beam, a 501-billion-parameter open model taking on China's giants

Reflection AI has unveiled Beam, an open-source large language model (LLM) with 501 billion parameters. The launch comes just months after the startup closed a funding round that valued it at $25 billion. According to information cited in the article, around the same time Reflection AI reportedly signed a $6.3 billion deal with SpaceX Corp. to rent Nvidia GB300 NVL72 systems, each equipped with 72 GPUs — infrastructure the company used precisely to train Beam.

In its own benchmark tests, Reflection AI measured Beam against GLM-5.2, an open model with roughly 250 billion more parameters. The company claims Beam matches or beats that rival on certain tasks while using between a third and a quarter of the hardware required. It also maintains that its model approaches the performance of Qwen 3.8-Max, a system with more than 2 trillion parameters. These figures come directly from comparisons run by the company itself, so they are best treated as the starting point of its sales pitch rather than as independent validation.

The launch matters because of the context in which it arrives: many of the most advanced open models available today —among them Qwen 3.8-Max and GLM-5.2— come from Chinese companies. Beam is thus presented as the first open-source model from a US startup to demonstrate performance comparable to or better than those alternatives, making it a symbolic milestone for the Western open ecosystem. Even so, the article itself notes that free models still lag behind closed frontier systems, citing Anthropic's Claude Fable 5.1 as an example.

As for the development process, Reflection AI explains that it began by training a relatively small prototype and progressively scaled up towards larger and more capable models, arriving at Beam Base, the foundation on which Beam is built. This base model was developed on a 6,144-GPU cluster and trained on 23.8 trillion tokens from the public web and commercial sources, including a significant amount of software code; to clean the latter, the company created custom filters for each programming language in order to remove low-quality files. The entire process was completed in under four weeks.

That phase was followed by so-called "midtraining", an optimization process aimed at extending the model's context window and improving its reasoning ability, which served as preparation for the third and most demanding training stage. For it, Reflection AI deployed 10,000 GB300 GPUs to launch 1.3 billion reinforcement learning environments (sandboxes), virtual spaces in which the model learns new skills. These environments were optimized specifically for code generation, web search and AI agent execution tasks, and this phase was also completed in four weeks.

The company also highlights the engineering effort devoted to keeping training running without interruption: it developed software to prevent failures from breaking the workflow, achieving a median recovery time of eight minutes against the 71 errors recorded over the whole process. It is a figure that illustrates the operational scale required to train models of this magnitude and the growing importance of infrastructure resilience as a competitive advantage.

For now, Beam is available only through an early access program. Reflection AI has announced that it will release the model's weights, its documentation and fine-tuning tools later this month, which will allow the community to independently verify the performance benchmarks the company alone has presented so far.