Unverified50% confidenceReleaseExact time
Xiaomi additionally shared a separate MTP model.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/24/2026
First Seen
Valid until: 10/22/2026
Sources
小米静默发布MiMo-V2.5-DFlash：300B模型推理速度或将翻倍
redditr/LocalLLaMA7/12/2026
Related Claims
UnverifiedMTP一次前向传播可同时产出多个token：主头输出token+1，第一个MTP头输出token+2，第二个MTP头输出token+365% similarUnverifiedllama.cpp has officially incorporated MTP (Multi-Token Prediction) technology into mainline support.61% similarUnverifiedCopilot's Sampling mechanism supports configuring the allowed models separately at the global level and at the individual MCP Server level.56% similarUnverifiedZotero MCP插件通常采用stdio模式，数据在进程间直接传递，不经过本地网络56% similarUnverifiedCommunity members implemented MTP support on a fork of llama.cpp (ik_llama.cpp) and conducted real-world testing based on the Qwen 3.6 model.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/607368API
curl https://kongchang.com/api/v1/knowledge/claims/607368MCP
get_claim(id=607368)