Reflection AI Open-Sources Beam Weights and Unveils Training Recipes

American company Reflection AI has open-sourced the weights of its language model, Beam, to everyone on the Hugging Face platform, releasing a suite of technical resources that allow developers to customize and run the model.

The company had unveiled Beam a week ago, featuring approximately 500 billion parameters. In some benchmarks, its performance approaches that of models four times its size while consuming less compute power and memory. The company states it trained the model over two months using several thousand advanced GPUs.

To address training errors at scale, Reflection AI developed proprietary reliability algorithms focused on avoiding mismatches between training and inference, an issue that occurs when configuration settings differ across training environment components. It also built a system that automatically fixes data quality errors and rollbacks to a previous checkpoint instead of starting over, enabling training with 'no actual cluster downtime,' as the company described it.

The available releases include a standard version and a smaller quantized version that reduces representation precision in exchange for higher hardware efficiency. The package includes training recipes, customization code templates, tools to measure the impact of modifications on capabilities and safety guardrails, a 'draft' model to accelerate inference via speculative decoding, and ready-to-use configurations for popular serving engines. The company plans to publish additional recipes in the coming days that will allow the model to refine its outputs through self-distillation.

What do these terms mean?

Model weights: The numerical values representing what the model has learned; publishing them allows running or retraining the model outside the company's infrastructure.

Quantization: Compressing the model's numbers to store them in less space, reducing memory usage and cost at the expense of a slight decrease in accuracy.

Speculative decoding: A technique that uses a small model to generate a fast draft that the larger model verifies, thereby reducing latency.

Self-distillation: Training the model to improve its outputs based on a refined version of its previous answers rather than relying solely on external data.