MTP 技术让本地大模型推理速度翻倍，尤其适合在消费级显卡上跑密集模型的开发者——2 块 RTX 5090 就能让 27B 模型达到 117 tps，值得直接试开源代码。
Atomic Chat 团队通过 Multi-Token Prediction (MTP) 技术，在 2 块 RTX 5090 上对 Qwen 模型实现了最高 2.5 倍的推理加速。其中，Qwen3.6 27B 密集模型从 51 tps 提升至 117 tps（+137%），而 MoE 模型 35B-A3B 从 218 tps 提升至 267 tps（+25%）。MTP 通过一次前向传播验证多个预测 token，显著减少了内存带宽瓶颈，密集模型受益更大。该技术保持零精度损失，仅需额外约 1 GB 显存，且代码已开源。
MTP speedup Qwen by 2.5x in Atomic Chat Dense vs …
MTP speedup Qwen by 2.5x in Atomic Chat
Dense vs MoE models on 2x RTX 5090 Qwen3.6 27B: 51 → 117 tps +137% Qwen3.6 35B-A3B: 218 → 267 tps +25%
MTP drafts several tokens ahead and verifies them in one pass. The speedup depends on memory moved per pass. Dense 27B reads all 27B params per token, MoE 35B-A3B only reads 3B active. Dense had way more to save by batching.
The baseline tps also differ (218 vs 51) for the same reason from the other side. Token generation is memory-bandwidth bound, and MoE moves ~8x less memory per token, so its baseline is already 4x ahead.
~80% draft acceptance. Zero accuracy loss. ~1 GB extra VRAM.
Open-source code and local AI app – in the comments 👇