Reflection releases open Beam 1.0, a 501B MoE model
Original title: reflection/Beam-1.0-501B-A23B — new model trending #30 on Hugging Face
The instruction-tuned text model activates 23.4B of 501.3B parameters and supports a 1M-token context.
Reflection published Beam-1.0-501B-A23B, an Apache 2.0 instruction-tuned text model in BF16, FP8, and NVFP4. It is a sparse mixture-of-experts model with 501.3B total parameters, 23.4B active parameters, 256 experts, and a 1M-token context, developed predominantly for English. The card claims token-efficient reasoning, coding, agentic tool use, and math and science reasoning; a technical report is listed as coming.
Unconfirmed: Capability claims are Reflection’s model-card statements; the technical report is not yet published.
- Architecture is 52 layers: 3 dense and 49 MoE, with gated GQA.
- Vocabulary size is 152k tokens; the chat template is a modified Harmony format.
- Knowledge cutoff stated on the card is June 30, 2026.
- The card claims about 3x fewer reasoning tokens than leading open models.
- Vendors
- Reflection
- AI models
- Beam-1.0-501B-A23B
From the original article
5 min read at huggingface.cotext-generation · 2,097 downloads · 163 likes · license apache-2. 0
By reflection · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.