A.I.
Rycode

Jeffrey 1 min read

Open weight · new lab

Beam 1.0 (Reflection)

Beam 1.0 is the first model from Reflection, a US lab. It was released on 16 Oct 2026 as open weights, meaning anyone can download the model files and run them. The license is Apache 2.0, which allows commercial use. Reflection says Beam is built for coding and long agent tasks, where the model works through many steps on its own.

Status
Current, open weights, Apache 2.0
AA score
30 (AA, 16 Oct 2026, see note 1)Artificial Analysis Intelligence Index, one independent score across many tests, fetched 16 Oct 2026
Released
16 Oct 2026
Last checked
16 Oct 2026newest source date on this page

Artificial Analysis (AA), an independent group that tests models, ranks Beam second among US open-weight models, with 30 on its Intelligence Index (16 Oct 2026 ET). Only Meta's Muse Spark Open scores higher among US open-weight models, per AA.

What is published

From Reflection's launch post and the Hugging Face launch blog:

  • Mixture of experts, a design where only part of the model runs for each token (a token is a word or piece of a word)
  • 501B parameters in total, 23B active per token
  • 1M-token context window, the amount of text it can read at once
  • Text in, text out
  • Weights in BF16, FP8 and NVFP4 formats (different number precisions that trade size for accuracy)

Benchmarking

AA Intelligence Index: 30, second among US open-weight models (AA article, 16 Oct 2026 ET). Reflection publishes its own lab test results on its model card. Those are Reflection's numbers and are not copied here. Reflection lists AA among the partners that tested Beam before release.

Access

Reflection says Beam is available at launch on Ollama, Together AI, Modal and Cloudflare, among others, and runs on vLLM, an open-source serving library.

Sources