Introducing Beam: Reflection’s 501B open-weight model

Beam is a 501-billion-parameter sparse Mixture-of-Experts model released with open weights under Apache 2.0 and documented with a model card and a comprehensive safety technical report. Designed for advanced coding, long-horizon reasoning, and agentic workloads, it activates roughly 23 billion parameters per pass and ships with benchmarks and safety evaluations. Reflection positioned Beam as the kernel of an "open operating system for agents," launching with 40+ partners and integrations so developers can self-host, run managed endpoints, or connect via Reflection’s API. Early demos show complex agentic use cases - interactive puzzle solving, generative voxel scenes, and rapid game-building - illustrating token efficiency and real-world latency gains.

Deployment tooling and customization are central: weights are available in BF16, FP8, and NVFP4, plus a speculative-decoding drafter, with partner-optimized serving kernels to lower memory and speed up generation. Beam integrates natively with transformers, vLLM, SGLang, AXOLOTL and others, and ships with Mirror CLI, an open-source terminal harness whose Python frontend and Rust core support resumable sessions, parallel agents, tool approvals, and inspection. Reflection provides recipes for fine-tuning, large-scale RL (1.3 billion sandbox runs during training, peak fleet at ~883k vCPUs), and upcoming self-distillation workflows. Beam is available across major cloud, inference, and on-prem platforms, with documentation, weights on Hugging Face, and community channels for developers.