Reflection AI open-sources Beam weights and post-training recipes

Reflection AI has published the weights of Beam, its first large language model with about 500 billion parameters, on Hugging Face along with fine-tuning recipes, evaluation tools and a drafter model for speculative decoding. The company trained Beam over two months on several thousand high-end GPUs with near-zero cluster downtime.

• Beam has about 500 billion parameters and rivals models four times larger on some benchmarks
• Training took two months on several thousand high-end GPUs
• Standard and quantized weights, recipes and a drafter model are available
• Ready-to-use configs for popular serving engines and fused kernels included