Reflection AI Introduces 501B Parameter Open-Source Beam Model
Reflection AI has unveiled Beam, a 501-billion-parameter open-source language model trained on Nvidia GB300 hardware that rivals top Chinese open-source systems.
Reflection AI Inc. has introduced Beam, an open-source large language model built with 501 billion parameters. The announcement comes shortly after the startup secured funding at a $25 billion valuation. The company also reportedly inked a $6.3 billion agreement with SpaceX Corp. to rent Nvidia GB300 NVL72 systems, each housing 72 graphics cards, which were utilized to train the new model.
Training pipeline and efficiency benchmarks
Reflection AI initiated the project by training a smaller prototype before progressing through a sequence of increasingly capable models to build Beam Base. This foundation model was trained over a period of less than four weeks using a cluster of 6,144 graphics cards. The training dataset comprised 23.8 trillion tokens from public web pages and commercial sources, featuring a significant volume of software code that was refined using custom filters developed for individual programming languages.
Following initial training, Reflection AI conducted a midtraining phase designed to broaden the model's context window and strengthen its reasoning features. For the final reinforcement learning stage, the startup deployed 10,000 GB300 graphics cards to launch 1.3 billion virtual sandboxes. These environments focused on specialized tasks, including running AI agents, searching the web, and generating code. The reinforcement learning phase was completed in four weeks, supported by custom fault-recovery software that maintained a median recovery time of eight minutes across 71 hardware and system errors.
In evaluation tests against GLM-5.2—an open-source model containing roughly 250 billion additional parameters—Beam delivered superior performance on select tasks while using one-third to one-fourth of the hardware compute. Reflection AI also stated that Beam approaches the performance of Qwen 3.8-Max, a model featuring more than 2 trillion parameters. Although open-source architectures continue to trail frontier commercial models like Anthropic PBC's Claude Fable 5.1, Beam represents the first open-source release from a U.S. startup to match or exceed the performance of top Chinese open-source models like Qwen and GLM.
What it means for developers
Beam provides developers with an open-source option capable of complex reasoning, web searching, and software code generation without requiring the massive parameter scale of multi-trillion-parameter alternatives. The optimization techniques used during training demonstrate that smaller, carefully filtered datasets and targeted reinforcement learning sandboxes can yield high task performance with lower compute requirements.
While Beam is currently limited to an early access program, Reflection AI plans to make the model weights, technical documentation, and fine-tuning tools publicly available later this month. Once released, engineering teams will be able to customize and host the 501B-parameter architecture for custom agent workflows and programming tools.
For teams evaluating different model architectures for their application stack, developers can try top AI models cheaply through one API at https://apixoai.online.
Release schedule and ecosystem impact
The upcoming release of Beam's weights and fine-tuning tools marks a significant entry into an open-source ecosystem recently dominated by Chinese tech firms. By sharing its full model assets and documentation, Reflection AI offers developers and researchers a powerful U.S.-developed foundation model optimized for multi-step tasks, agent execution, and rapid error recovery in distributed clusters.
Source: Reflection AI debuts open-source Beam model with 501B parameters — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
