X · @teortaxesTex
· X / Twitter
If they ignore latency, they can do like 18K tokens/second/GPU of ≈Opus 4.7 "Flash", anon Just on two SuperPoDs they had by May, that's 295 million t…
If they ignore latency, they can do like 18K tokens/second/GPU of ≈Opus 4.7 "Flash", anonJust on two SuperPoDs they had by May, that's 295 million tokens/second. 25.5 trillion tokens a day. That's a lot of trajectories.In case you wondered how they can pull off these RL feats.Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞): In t