The developer documentation says each recipe includes the configurations and scripts used by Reflection, hardware and runtime records, and, where training targets quality, before-and-after measurements made with the same evaluation method.
The listed work includes Beam self-distillation on verified Triton kernels; SWE-agent and tool-calling-reasoning distillation into Gemma 4 E4B; and GRPO reinforcement learning on DAPO-Math-17K through Miles, NeMo RL, and SkyRL. The page says one data-synthesis recipe has Beam generate tool-calling thoughts for 3,941 public conversations. Another creates software-engineering environments with hidden tests, retains only passing trajectories, and uses them for training data.
Reflection also gives sizing guidance for Beam-1.0-501B-A23B, described as 501B total parameters with 23B active. It lists one 8x H200 node for QLoRA, three to eight nodes for full fine-tuning, and four to eight for RL post-training. The page does not establish that these recipes will reproduce the reported results on other tasks, datasets, hardware, or deployments.