Unverified85% confidenceFactTime unknown
OmniMerge V4是由三到四个27B模型通过权重混合融合而成的模型
1
Sources
85%
Confidence
Medium-term (~90 days)
Relevance
6/1/2026
First Seen
Valid until: 8/30/2026
Sources
Qwen3.6 27B三大邪修量化模型实测：代码暴增15.8PP、40B蒸馏、16GB适配
bilibili尚书省说书人
Related Entities
Related Claims
VerifiedINT4量化理论上可将模型显存占用压缩至FP16的四分之一，常用工具包括GPTQ、AWQ、GGUF等70% similarUnverified基础模型（无LoRA）配置为30步、CFG 4、er_sde采样器、simple调度器70% similarUnverifiedOmniMerge V4的Q5版本可以支持100K的上下文长度66% similarUnverifiedOllama默认提供的模型多为4-bit量化版本（Q4_0或Q4_K_M格式）65% similarUnverifiedRTX 30系列（Ampere架构）的Tensor Core对INT4的原生支持有限，INT4运算通常需要拆包模拟65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18008API
curl https://kongchang.com/api/v1/knowledge/claims/18008MCP
get_claim(id=18008)