Reflection announced Beam’s open-weight release on Oct. 16 under the Apache 2.0 license, saying the model is available across platforms through an ecosystem of more than 40 partners. The release includes model weights, a tokenizer and a reference implementation, according to the company.
Reflection also lists BF16, FP8 and NVFP4 weights, a DFlash2 drafter and Mirror CLI, the open-source harness it says it built and trained Beam on. The model and its accompanying harness are being released together, rather than as a weights-only announcement.
Artificial Analysis describes Beam as a text model with 501 billion total parameters, 23 billion active parameters and a 1 million-token context window. In its Oct. 16 evaluation post, the organization reports an Intelligence Index score of 30, second among the U.S. open-weight models it evaluated, behind Muse Spark Open’s 39. It reports 48.5% on AutomationBench-AA but 3.0% on Terminal-Bench 4.0.
The evaluator separately reports a Coding Agent Index score of 28 for Beam running in Mirror. That places the setup below the next entry at 43 and the leaders at 68, even as Artificial Analysis highlights its low token use.
