Qwen 3.5 hybrid attention triggers long-context hallucinations

The transition to hybrid attention for inference efficiency is exposing a painful trade-off between raw context length and generation stability.

• –While Qwen 3.5's mix of linear and full attention enables massive 256k windows on paper, users report the model struggles to maintain coherence in complex, real-world prompts
• –The issue highlights a broader industry challenge where theoretical "needle in a haystack" benchmark success doesn't always translate to reliable agentic workflows
• –Developers are currently finding that standard mitigation techniques are ineffective against these underlying architectural quirks
• –This may drive users prioritizing factual integrity in long documents back to computationally heavier, full-attention models