RTX 3090 hits 4.6x faster token generation on a 27B LLM via inference optimizations

A developer running Qwen3.8-27B on an RTX 3090 raised token generation from 82 to 382 tok/s on a 25,000-token document task. The gains came from speculative decoding, MTP and a lookup drafter; ordinary chat runs at about 133 tok/s with math accuracy steady at 96.5%.

• Speed rose from 82 to 382 tok/s on a 25,000-token document task
• MTP lifted generation to 114 tok/s, DFlash2 and lookup to 138 tok/s
• The engine checks 16 tokens per step, accepting about 15 on quotes
• Context window extended to 138,696 tokens via compact memory format