26.4B model claimed to beat 975B: 17x compute efficiency gain in 128 days
teortaxesTex · x · 2026-11-21
teortaxesTex cites TML data: since July 15, a 26.4B total / 3.4B active (5.4B effective) model surpasses a 975B 41BAB model — a geometric-mean compute efficiency gain of 17x, possibly 50x on pretraining efficiency, achieved in 128 days. He argues an Opus-5.5-class cognitive core can run on a single consumer GPU, noting part of the gain is one-time but such optimizations are repeatable. Unverified third-party analysis.
More from Infra
- US chip crackdown on Southeast Asia hands Ambani and Adani a $200B AI infrastructure windfall — nordicinst · 2026-11-21
- Nvidia's $2.34T market cap now rivals Germany's entire stock market — HanchungLee · 2026-11-21
- Chollet: AI made staggering leaps since 2024, yet NVDA buyers are down 5% — fchollet · 2026-11-21
- Meta rethinks 2030 net-zero goal as AI data centers drive record emissions growth — Ars Technica AI · 2026-11-21
- Oracle says it's fully funded for OpenAI contract, first hyperscaler to complete this cycle's raise — Scobleizer · 2026-11-21
- NVIDIA posts NVFP4 checkpoints tuned for DGX Spark as Blackwell runs FP4 natively — AIFlow_ML · 2026-11-21