Unverified60% confidenceFactTime unknown
RTX 5090上Qwen3.6 27B模型Q4量化不开MTP解码速度约63 token/s，开启MTP后跳到105 token/s
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
6/1/2026
First Seen
Valid until: 8/30/2026
Sources
Qwen3.6 MTP加速实测：单GPU推理飙到220 token/s
bilibili普通鱼学家呀
Related Entities
Related Claims
UnverifiedQwen3.6 35B模型借助MTP技术在oMLX上达到了86.7 tokens/s的生成速度，prompt处理速度为1735 tokens/s85% similarUnverified千问3.6 27B FP8模型开启MTP后正常速度约38 token/s，平均测速可达70 token/s以上78% similarUnverified在显存充足、模型完整驻留GPU时，RTX 4090跑Q4量化的27B模型通常可达15～30 tokens/s；触发CPU卸载后速度可能骤降至1～2 tokens/s甚至更低76% similarUnverifiedRTX 5090上Qwen3.6 27B模型预填充速度从174提升到253 token/s，显存占用约18GB74% similarUnverified在24GB显存的显卡上运行Qwen3.6 27B模型，未经优化可达每秒约40个Token的生成速度，优化后可达50-60 Token/s74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18058API
curl https://kongchang.com/api/v1/knowledge/claims/18058MCP
get_claim(id=18058)