Reflection AI open-sources Beam LLM weights and recipes

Reflection AI has released the weights of Beam, its 500-billion-parameter large language model, on Hugging Face, along with post-training recipes, fine-tuning tools, and serving engine configurations. Beam approaches the performance of models four times its size on some benchmarks while using less compute and memory. The release includes quantized variants, a drafter model for speculative decoding, and fused kernels for faster inference. Reflection AI developed custom reliability algorithms during training to minimize downtime and avoid training-inference mismatches. The company plans to release additional self-distillation training recipes and open-source contributions in the coming days.