Expired50% confidenceBenchmarkExact time
After a developer applied abliteration to Qwen3-VL-4B-Instruct, the HarmBench compliance rate reached 100%, compared to only 30.8% for the original model.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/20/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/18/2026(expired)
Sources
Related Claims
UnverifiedThe original Qwen3.6 27B Q4 scored 97 in the tool-calling test, demonstrating that distillation does not always improve performance.75% similarUnverifiedQwen3.6 27B NVFP4 achieved a perfect score of 100 in the tool-calling (ToolCore) test.71% similarUnverifiedThe Qwen3.6 34B 80-layer model achieves an MTP speculative decoding acceptance rate of approximately 78% on LM Studio, while the original 27B achieves only approximately 57–58%70% similarUnverifiedThe Qwen3.6 35B model contains 256 experts, with each token activating only 8 routed experts and 1 shared expert, resulting in approximately 3B activated parameters during actual inference.69% similarUnverifiedThe Qwen3.6 27B version achieved an accuracy of 95.7% on the Simple QA test.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)