Qwen3.5 stumbles on multi-system chat prompts

This is the kind of annoying integration bug that matters more than benchmark wins when people actually try to use local models in real apps.

• –The core problem is API-shape compatibility: roleplay presets that rely on multiple `system` messages do not map cleanly onto every OpenAI-style backend.
• –Qwen’s official ecosystem leans on chat templates and structured message handling, so text completion working while chat completion breaks is consistent with a formatting mismatch, not a raw capability gap.
• –For local AI tooling, this is a reminder that SillyTavern presets, llama.cpp server behavior, and model-specific templates can be just as important as the model weights themselves.
• –The workaround implied by the thread—merging system prompts or rewriting the Jinja template—keeps things running, but it can also subtly change behavior in roleplay-heavy setups.
• –This is useful signal for self-hosters: “OpenAI-compatible” still often means “compatible enough, with caveats,” especially for multi-message system prompting.