Unverified50% confidenceFactExact time
分享者在大模型时代之前训练过的最大模型参数量仅约4000个参数
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Verified对于一个 70 亿参数的模型，LoRA 通常只需训练不到 1% 的参数量，显存需求可降低 60% 以上65% similarUnverified1000亿参数的模型处理单次推理请求需约80GB显存，百亿参数级别模型仅需约20GB65% similarUnverified分析推测该模型是2到4万亿参数规模的MoE模型，每个Token仅激活约1500亿参数65% similarUnverified早期LLM只能处理约4000个token，当前最新模型已支持10万甚至100万token的上下文64% similarUnverified参数量在1亿以下的轻量模型经领域数据微调后可达95%以上意图分类准确率，单次推理延迟控制在5-20ms以内64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/908993API
curl https://kongchang.com/api/v1/knowledge/claims/908993MCP
get_claim(id=908993)