Unverified50% confidenceFactExact time
MOSS-VL-Realtime提供4比特量化版本，按单张RTX 4090设计，110亿参数模型在4比特量化后权重占用可降至约5.5GB显存
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/10/2026
First Seen
Valid until: 12/9/2026
Sources
Related Entities
Related Claims
VerifiedQLoRA引入4-bit量化技术，使得在单张24GB显存的RTX 3090/4090上微调650亿参数级别的模型成为可能74% similarVerified700亿参数（70B）的模型经量化后可以在单张RTX 4090（显存24GB）上运行69% similarUnverified约25GB的Q3模型完全在RTX 4090显存（带宽约1TB/s）运行理论上每秒约40个token，而在DDR5内存（约80GB/s）中运行降至每秒1-3个token69% similarUnverified量化后70亿参数模型的显存需求从约28GB压缩至约4-6GB，普通消费级显卡如RTX 3060/4060即可流畅运行Llama 3、Qwen2.5、Mistral等70亿参数级别模型68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/885998API
curl https://kongchang.com/api/v1/knowledge/claims/885998MCP
get_claim(id=885998)