Reflection on October 16, 2026 released the weights for Beam, its first open-weight model, under an Apache 2.0 license. The company published the checkpoints on Hugging Face in BF16, FP8, and NVFP4 formats, alongside a speculative decoding drafter, a model card with final benchmarks, and a Beam Safety Technical Report.
Beam is a sparse mixture-of-experts model with 501.3 billion total parameters and 23.4 billion active per token, a 1 million-token context window, and text-only input and output. Reflection positions it for coding, reasoning, and long-horizon agentic workloads, and pairs the release with Mirror, an open-source CLI harness the lab trained alongside the model.
Independent evaluator Artificial Analysis scored Beam 30 on its Intelligence Index, placing it second among U.S. open-weight models behind Meta’s Muse Spark Open (39) and ahead of NVIDIA’s Nemotron 3.5 Super VL (26), all announced in the same week. Artificial Analysis also highlighted Beam’s token efficiency (~26k output tokens per Index task) and stronger AutomationBench-AA results relative to several open peers, while noting weaker Terminal-Bench 4.0 scores.
Reflection says Beam has day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and NVIDIA Dynamo, and lists managed or partner paths including AWS, Google Cloud, Microsoft Foundry, Fireworks AI, Together AI, OpenRouter, Ollama, and others. Developers can try Beam on platform.reflection.ai, pull weights from Hugging Face, or follow deployment guides at developers.reflection.ai.
DigiEditorial treats Reflection’s October 16 availability post as the primary source and corroborates the open-weight launch window, Apache 2.0 licensing, parameter counts, and Artificial Analysis ranking with Artificial Analysis’s own write-up and Hugging Face model listings.