Reflection's Beam matches DeepSeek V3 flops but burns 3.5-4x more hardware, low utilization
teortaxesTex · x · 2026-11-08
- Teortaxes dissects Reflection's training efficiency after ex-employee complaints about under-utilized GPUs.
- Inkling 2S: pretraining run is comparable to DeepSeek-Coder-V2 from June 2024 (21B active vs 17B, 10.2T vs 12.25T tokens) — yet it finished 12x faster, mostly due to hardware (bf16 vs fp4).
- Beam 1.0: an iso-flop replication of DeepSeek V3 (3.3e24 FLOPs via 23B×23.8T vs 37B×14.8T), but V3 used 2.66M H800-hours while Beam used 4.13M GB300-hours. Hardware-normalized, Beam burned 3.5–4x more raw capacity, with implied utilization in single digits to 10% vs V3's 35%. Alex Polozov pushed back citing 92% goodput.
- Takeaway: Inkling 2S is roughly at V4.1 Flash level, but the efficiency gap shows "neolabs are not made equal."
Related event: Ex-Employee Exposes Google-Style Bureaucracy at Reflection AI(4 posts)→
More from Infra
- Zyphra's ZHA sparse attention keeps 35B model quality while reading just 1-6% of KV-cache — burny_tech · 2026-11-08
- llama.cpp merges PR #30966: run missed MoE experts on CPU instead of PCIe transfers — QuixiAI · 2026-11-08
- DGX Spark prices are starting to come down — AIFlow_ML · 2026-11-08
- MOSS-Image 5B runs fully on a phone NPU: ~30s for 512x768 on Snapdragon 8+ Gen 1 — shirosoro · 2026-11-08
- Open-weight models now handle 72% of OpenRouter tokens: build your fallback before prices bite — shashib · 2026-11-08
- CoreWeave CEO: one data center uses about as much water yearly as one Krispy Kreme store — aronchick · 2026-11-08