Reflection AI Open-Sources Weights and Post-Training Recipes for 501B-Parameter Coding Model Beam

AI startup Reflection AI released the weights of Beam, its first large language model, on Hugging Face on October 16 under an Apache 2.0 license. Beam is a sparse mixture-of-experts (MoE) model with approximately 501 billion total parameters and 23 billion active parameters per token, specifically designed for agentic coding tasks such as resolving software issues, modifying code, and executing multi-step workflows.

The release includes not only the model weights but also complete post-training recipes, code templates, and a lightweight drafter model for speculative decoding. These resources allow developers to fine-tune the model to enhance programming capabilities in specific domains or adjust safety guardrails. Additionally, Reflection AI provided ready-to-use configurations and fused kernels for popular inference engines like vLLM and SGLang to lower deployment barriers.

According to Reflection AI, Beam was pretrained on 23.8 trillion tokens and optimized through reinforcement learning involving over 100 million rollouts on 10.5K NVIDIA GB300 GPUs. The company reports that Beam achieves high scores on benchmarks like GPQA Diamond and SWE-bench Verified, approaching the performance of models four times its size while using significantly less computing power and memory. These benchmark results have not yet been independently reproduced.