Reflection releases Beam model weights on Hugging Face under an Apache 2.0 license | Model Release
On October 16, 2026, Reflection released Beam's model weights on Hugging Face under an Apache 2.0 license. Beam is a sparse Mixture of Experts model with 501 b…
Published on MyPrivateClaw
Oct 16, 2026, 11:37 PM UTC
Coverage date
Oct 16, 2026
Last updated
Oct 16, 2026, 11:37 PM UTC
News summary
On October 16, 2026, Reflection released Beam's model weights on Hugging Face under an Apache 2.0 license. Beam is a sparse Mixture of Experts model. It carries 501 billion total parameters, with 23 billion active parameters used per token for coding, reasoning, and agentic workloads, per the vendor's active parameter efficiency claim. At launch, the weights were released in BF16, FP8, and NVFP4 precision formats alongside a speculative decoding drafter model, with day zero support for Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and Dynamo inference runtimes. Why it matters Beam's significance in this release is its packaging and distribution rather than a benchmark rank. On October 16, 2026, the weights became downloadable under an Apache 2.0 license, which is what lets an operator host the model in its own stack instead of only reaching it through a vendor endpoint. The…