🤖Reflection Beam: 501B MoE With 23B Active, Weights Due
TL;DR
Reflection announced Beam, a text-only mixture-of-experts model with 501B total and 23B active parameters, trained from scratch on about 23.8T tokens. It claims 80.9 on SWE-bench Verified, self-reported.
Reflection announced Beam, a text-only mixture-of-experts model with 501B total and 23B active parameters, trained from scratch on about 23.8T tokens. It claims 80.9 on SWE-bench Verified, self-reported. Apache 2.0 weights are promised this month.
Key Points
501B total, 23B active parameters; built for coding, agentic and scientific work
Self-reported 80.9 on SWE-bench Verified and 3-4x the inference efficiency of GLM 5.2
Independent observers place it near GLM-5.2, below DeepSeek V4 Flash on some tests
Full Apache 2.0 weights due this month; a tech report is promised
Axios reportedly says Reflection pays $150M/month for Colossus compute plus a $1B Nebius deal
Why It Matters
It is the first serious US open-weight entrant at this scale in a while, but weights and independent evals decide whether it matters. Wait for both.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,564 builders reading daily.