Conclusion
Qwen 3.6 Max Preview scores 94.5 on the BridgeBench BS Benchmark (anti-hallucination/nonsense detection test), ranking second globally. This benchmark specifically tests whether models can identify and refuse to generate false information when faced with leading questions.
Rankings:
- Claude Opus 4.6: 95.0
- Qwen 3.6 Max: 94.5
- Claude Sonnet 4.6: 91.5
- GPT-5.4: 91.5
Qwen 3.6 Max is the highest-ranking open-source model and the only one whose anti-hallucination capability exceeds all OpenAI models among non-closed-source options.
Test Dimensions
What Is the BS Benchmark?
The BS Benchmark (Bullshit Benchmark) tests a core capability: when users ask questions containing false premises, misinformation, or logical traps, can the model identify the problem itself rather than blindly generating plausible but actually wrong answers?
This differs from traditional knowledge tests — traditional tests ask “what do you know,” while the BS Benchmark asks “do you know what you don’t know.”
Qwen 3.6 Max Performance
Qwen 3.6 Max’s score of 94.5 means that in the vast majority of test scenarios, it can:
- Identify false premises in questions and point them out
- Express reasonable doubt when uncertain rather than fabricating answers
- Distinguish between “well-founded speculation” and “baseless guessing”
Notably, Qwen 3.6 Max scored higher than GPT-5.4 (91.5) and Claude Sonnet 4.6 (91.5), trailing Claude Opus 4.6 by only 0.5 points.
Significance for the Open-Source Ecosystem
For a long time, anti-hallucination capability was considered the “moat” of closed-source models. Qwen 3.6 Max’s performance proves that open-source models have caught up and in some aspects surpassed closed-source alternatives on this critical metric.
For scenarios requiring high-reliability output (healthcare, legal, finance), Qwen 3.6 Max provides an open-source alternative without vendor lock-in concerns.
Selection Guidance
- High-reliability scenarios: Qwen 3.6 Max’s anti-hallucination capability approaches top closed-source models, suitable for applications with strict output accuracy requirements
- Open-source-first strategy: If your team needs self-hosting or wants to avoid vendor lock-in, Qwen 3.6 Max is currently the strongest open-source choice for anti-hallucination
- Cost considerations: Open-source deployment avoids per-token API costs, especially valuable for high-volume scenarios
- Multi-model collaboration: Use Qwen 3.6 Max as a fact-checking layer alongside other models that generate content