Reflection Beam 1.0 weights: 501B MoE, 23B active, Apache 2.0

Reflection published the weights for Beam 1.0 on October 16, 2026, in a post titled "Beam is now available". It announced the model on October 5, 2026. Beam is a sparse mixture-of-experts model with 501B total and 23B active parameters. Reflection's Hugging Face model card lists 501.3B total and 23.4B active parameters, 52 layers (3 dense, 49 MoE), 256 experts, a 1M-token context, and text-only input and output. The licence is Apache 2.0, and Reflection's Hugging Face repos are not gated. Reflection released BF16, FP8 and NVFP4 checkpoints plus a speculative-decoding drafter. The drafter repo, `Beam-1.0-501B-A23B-DFlash2`, lists 1.93B parameters. None of Reflection's four Hugging Face repos is a GGUF or MLX build. Reflection's post lists day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo. Its list of launch platforms also includes LM Studio and Ollama. As of October 20, Hugging Face listed six third-party quantizations of Beam. By repo creation date, they are GGUF builds from unsloth, AtomicChat and DevQuasar (October 16), an EXL3 3bpw build from diffbot that its card says is for two RTX PRO 6000 Blackwell cards (October 18), an MLX build from OpensourceWTF (October 19), and a VQ-GGUF build from Baekpica (October 20).