Unverified50% confidenceFactExact time
Mixtral 8x7B拥有约470亿总参数，但每次推理仅激活约130亿参数，推理成本接近13B稠密模型
1
Sources
50%
Confidence
Long-term
Relevance
8/26/2026
First Seen
Sources
Qwen模型登顶HuggingFace热度榜前四背后的深层逻辑
Related Entities
Related Claims
Unverified通义零码的MoE架构激活比例约为7.3%（350亿/4800亿），推理成本接近一个350亿参数的密集模型70% similarUnverified论文估算解码一万条推理链（每条约1.2万token）成本仅约720美元，可用此价格窃取前沿模型推理能力用于蒸馏66% similarUnverifiedFine-tuning a 7-billion-parameter model can be done on a consumer-grade GPU, potentially at a cost of only a few hundred dollars64% similarUnverified在标准云服务器上，一个3.8B参数模型的单次推理成本约为70B模型的1/2061% similarUnverifiedThe API call cost per million tokens for Anthropic's top-tier Claude series models may be several times to more than ten times that of lightweight models.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/802484API
curl https://kongchang.com/api/v1/knowledge/claims/802484MCP
get_claim(id=802484)