Unverified90% confidenceFactTime unknown
V4-Pro的总参数与活跃参数比例约为32:1的稀疏比
1
Sources
90%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
DeepSeek-V4发布：1.6T参数+百万上下文，开源模型新标杆
twitterdeepseek_ai
Related Entities
Related Claims
UnverifiedW4A4量化将权重和激活值均压缩至4位整数，理论上相比FP16可将显存占用降低至约1/469% similarUnverified据视频对比数据，V4 Pro在智能水平上与GLM 5.2相当且价格略低67% similarUnverifiedMiniMax M3 的激活参数与总参数比例 23B/428B 约为 5.4%67% similarVerifiedQ4量化将每个权重参数从16位压缩到4位，理论上可将模型体积缩小约75%65% similarUnverifiedMoE模型中'35B'指总参数量，每个token只激活少量Expert（通常2-4个），实际计算量可能仅相当于6-7B的Dense模型65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/27893API
curl https://kongchang.com/api/v1/knowledge/claims/27893MCP
get_claim(id=27893)