[AINews] Reflection Beam - 501B-A23B American Open Model
Reflection announced Beam, a 501B-total, 23B-active open-weight MoE model trained from scratch in the US.
The model is text-only and aimed at coding, agentic, and scientific work, with Apache 2.0 weights promised this month. Latent Space says Reflection claims 80.9 on SWE-bench Verified and 3–4x the inference efficiency of GLM 5.2. The note places Beam around GLM-5.2 level, behind newer Chinese frontier open models, but significant as a US-trained open option. Source: Latent Space's note
The model is text-only and aimed at coding, agentic, and scientific work, with Apache 2.0 weights promised this month. Latent Space says Reflection claims 80.9 on SWE-bench Verified and 3–4x the inference efficiency of GLM 5.2. The note places Beam around GLM-5.2 level, behind newer Chinese frontier open models, but significant as a US-trained open option. Source: Latent Space's note
score 6