Unverified50% confidenceReleaseExact time
Unsloth推出的1-bit量化版本将27B参数级别的模型压缩至仅6-8GB
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/24/2026
First Seen
Valid until: 11/22/2026
Sources
Related Entities
Related Claims
Unverified原版QW3.6 27B模型文件体积约为54GB，Prismo通过数学压缩技术压缩至不到4GB，压缩比超过13倍，且声称不影响模型性能78% similarUnverifiedFP8量化可将7B模型权重从约14GB（FP16）压缩至约7GB76% similarUnverifiedQ3_K_M将每个权重压缩至约3bit，27B模型权重体积可降至约12～13GB76% similarVerified32B参数规模的模型在量化后通常可以压缩到约16-20GB显存占用，一张NVIDIA RTX 4090（24GB显存）就能运行75% similarUnverified动态GGUF的UD-IQ2_XXS档位将原始BF16模型体积从54.7GB压缩到约9GB，体积压缩率83.5%，准确率保留82.5%74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795639API
curl https://kongchang.com/api/v1/knowledge/claims/795639MCP
get_claim(id=795639)