Models New
Reflection makes Beam weights available under Apache 2.0, with uneven agent benchmark results
Published:
Reflection has made its Beam-1.0-501B-A23B weights publicly available on Hugging Face. The repository commit labeled "Beam is available" is timestamped October 16 at 12:34 CEST, after this morning's edition. The repository is public and ungated at verification. That commit is the timing anchor for this dispatch, not the model repository's earlier creation date.
The model card describes a text-only sparse mixture-of-experts model with 501 billion total parameters, 23 billion active parameters and a one-million-token context window. Instruction-tuned BF16, FP8 and NVFP4 releases are listed under Apache 2.0. This is a weights-availability development, not the first announcement of Beam: Reflection previewed it on October 5 and said the weights would follow later in the month.
In its October 16 evaluation, Artificial Analysis gives Beam an Intelligence Index score of 30, behind Meta's Muse Spark Open at 39. Its results are mixed: 48.5% on AutomationBench-AA but 3.0% on Terminal-Bench 4.0. The same evaluator reports 33% accuracy and a 47% hallucination rate on AA-Omniscience. These are results under specific evaluation setups, not a verdict that Beam is best for every agent workload.
Reflection's own model card reports different benchmarks, including Terminal Bench v2.1; those numbers are not directly comparable with Artificial Analysis's Terminal-Bench 4.0 result. The company's model-development technical report is still marked as forthcoming in the card, while a separate safety technical report is already accessible.
Caveat. We verified public artifacts, their provenance and reported benchmark results; we did not download the large weights or run the model. A repository commit is a directly observed publication milestone, not proof of the exact first instant anyone could obtain access. The October 5 preview is dated background within this story, not fresh news. The evaluated terminal benchmarks and harnesses differ, so their percentages must not be compared as like-for-like.
Sources