The weights are now on Hugging Face in BF16, FP8 and NVFP4 versions. Reflection says Beam is built for coding, reasoning and agent work and lists more than 20 platforms offering it.
Reflection said on Friday it has released the weights of Beam, the open model it unveiled earlier this month, on Hugging Face under an Apache 2.0 licence. According to the model card, Beam is a text-only sparse mixture-of-experts model with about 501 billion total parameters, about 23 billion active per token, and a 1 million-token context window; weights ship in BF16, FP8 and NVFP4 formats, with a drafter model for speculative decoding. Reflection says Beam was built for coding, reasoning and long-running agent tasks, and is publishing fine-tuning and reinforcement-learning recipes plus an open-source command-line agent harness called Mirror. The company lists cloud and inference partners including AWS, Google Cloud, Microsoft Foundry, Oracle, CoreWeave, Databricks and Together AI, with Dell and HPE for on-premises use. SiliconANGLE independently reported the weights release and training recipes.
Sources: Reflection · Hugging Face model card · SiliconANGLE