Unverified50% confidenceFactExact time
Each MTP head shares the hidden state representation of the main model and contains only a small feed-forward network and an output projection layer, with a parameter count typically less than 1% of the main model.
1
Sources
50%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe MTP head parameter count is typically less than 1% of the main model and is jointly pre-trained with the main model.78% similarUnverifiedMTP的额外预测头直接复用Transformer的隐藏状态表示，因此额外的计算开销极小66% similarUnverified马尔可夫头仅需记录前一个token状态即可推导下一token修正概率，参数量通常不超过原模型的0.1%61% similarUnverified价值网络仅在约5%–10%的回合中作为关键判定依据使用60% similarUnverifiedWith well-calibrated FP8 inference, the loss in model output quality can typically be kept within 1%.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/889917API
curl https://kongchang.com/api/v1/knowledge/claims/889917MCP
get_claim(id=889917)