Same Qwen 3.8 27B. Same RTX 5090.
76.3 → 220.4 tok/s.
Only the decoder changed. MTP drafts three future tokens; the main model verifies them, and accepted tokens advance together.
Acceptance: 85.1%. Significant—but workload-dependent.
https://t.co/N7jfRxo8ce
@RippleXrpie 🌻TIME TO PUMP
🚀To support XRP community, Brad Garlinghouse Ripple XRP CEO initiated XRP Giveaway
💰First come,first served: https://t.co/IbUD1K18Mp
📌Note: you can take a gift only once. Please hurry up!!🐪
@thebearablebull 🌻TIME TO PUMP
🚀To support XRP community, Brad Garlinghouse Ripple XRP CEO initiated XRP Giveaway
💰First come,first served: https://t.co/IbUD1K18Mp
📌Note: you can take a gift only once. Please hurry up!!🐪
@_Crypto_Barbie 🌻TIME TO PUMP
🚀To support XRP community, Brad Garlinghouse Ripple XRP CEO initiated XRP Giveaway
💰First come,first served: https://t.co/IbUD1K18Mp
📌Note: you can take a gift only once. Please hurry up!!🐪