Qwen3-Coder-Next GGUF stalls in Android Studio

This smells like a chat-template or backend parser mismatch, not the base model suddenly forgetting how to answer. The clue is in the server log: generation stops cleanly on an end-of-message token while the grammar still expects a trigger, which is classic runtime friction.

• –Unsloth’s Qwen3-Coder-Next docs explicitly steer users toward specific local-runtime settings, including llama.cpp-style serving and a non-thinking output mode, so template compatibility matters a lot here.
• –Similar Qwen family issues have shown up in other runtimes when tool-calling or chat templates drift out of sync, especially with GGUF builds and quantized variants.
• –Android Studio’s AI integration may be stricter than a plain chat UI, so a partially correct template can work for a few turns and then fail when the conversation state gets more complex.
• –The most likely fix is to verify the exact chat template, stop relying on a mismatched grammar wrapper, and compare against a known-good llama.cpp or server setup.
• –If Qwen3.5 works while Qwen3-Coder-Next does not, that points to a model/template combination problem rather than Android Studio itself.