Reflection’s Open-Weight Beam Model

This post details the integration of Reflection's Beam model, an open-weight Mixture-of-Experts model with a large context window, into CoreWeave Forge's Serverless Inference. It covers the model's development path from training using Weights & Biases Models and supervised fine-tuning with Axolotl on CoreWeave infrastructure, to its serving performance benchmarks on NVIDIA GB300 GPUs. The post also highlights the model's suitability for software engineering and long-horizon agentic tasks, and provides guidance on evaluating task success, tool-use reliability, completion time, and cost per successful task.