FrontierThe story, in brief
Reflection's Beam becomes the most capable open-weight model built outside China
Reflection's Beam activates just 23B of 501B parameters — and matches GLM-5.2 on reasoning while using a quarter the compute.
The KeyNews take
Why it matters
Open-weight efficiency challenge to Chinese frontier labs. Beam's sparse MoE strategy shows a Western path to capability-per-compute parity without raw parameter count; practitioners evaluating open-weight alternatives gain a concrete efficiency benchmark.
The key facts
5 to knowReflection Beam: 501B parameters, 23B active per token (sparse MoE)
Targets GLM-5.2 parity on coding and reasoning benchmarks
Uses 3-4x less compute than comparison baseline (GLM-5.2 implied)
Open-weight release
Frames efficiency-over-raw-performance strategy against Deepseek and Qwen
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: Reflection has released Beam, its first open-weight model. The mixture-of-experts system activates just 23 billion of its 501 billion parameters per token and aims to match GLM 5.2 on coding and reasoning while using three to four times less compute. The strategy against Chinese open-weight rivals…