Unverified50% confidenceBenchmarkExact time
开发者对Qwen3-VL-4B-Instruct执行abliteration处理后，HarmBench合规率达到100%，而原始模型仅为30.8%
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/20/2026
First Seen
Valid until: 10/18/2026
Sources
Related Claims
Unverified原版Qwen3.6 27B Q4在工具调用测试中获得97分，说明蒸馏并不总能提升性能75% similarUnverifiedQwen3.6 27B NVFP4在工具调用（ToolCore）测试中获得满分100分71% similarUnverifiedQwen3.6 34B 80层模型在LM Studio上的MTP推测解码接受率约为78%，而原版27B仅约57-58%70% similarUnverifiedQwen3.6 35B模型包含256个专家，每个Token仅激活8个路由专家和1个共享专家，实际推理时激活参数约为3B69% similarUnverifiedQwen3.6 27B版本在Simple QA测试中达到95.7%的准确率68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/568110API
curl https://kongchang.com/api/v1/knowledge/claims/568110MCP
get_claim(id=568110)