Beam Called an Iso-Flop Replication of DeepSeek V3: Same FLOPs Two Years Later, Far Worse Efficiency

mike64_t · x · 2026-10-06

teortaxesTex calculates that Reflection's Beam (23B×23.8T tokens) is an iso-flop replication of DeepSeek V3 (37B×14.8T), ratio ≈0.9996 — the same compute two years later, but not V3's efficiency: a gimped H100 in quant-savvy Chinese hands beats a GB300 in NVL72. mike64t cites it as exhibit A that Western labs can't pre-train.

Related event: Reflection's Beam Falls Short of Chinese Open Models, Called an Iso-flop DeepSeek V3 Replica(5 posts)→

Original post →

More from Infra

Infra channel →