Qwen3.5-9B stalls on coding tasks

Looks more like an integration mismatch than a raw model failure. Qwen3.5 is pitched as a thinking, agent-friendly model family, but local coding stacks can still make a 9B model look like it is stuck when the real issue is output limits, stop conditions, or overloaded prompts.

• –Ollama’s Qwen3.5 page lists a 256K context window and “thinking” support, so the early cutoff is more likely coming from max-token settings, timeouts, or the surrounding tool loop than from the model’s official context ceiling.
• –Heavy wrappers like OpenCode or Claude Code can drown a small local model in system prompt baggage, tool chatter, and planning overhead, which makes 9B variants brittle on multi-step coding jobs.
• –Community feedback in the thread points at parameter support and quantization details, which is a clue that model behavior can change a lot depending on the serving stack.
• –If the goal is reliable coding help, simpler prompts and lighter agent loops usually help more than just swapping between 4B and 9B.
• –For serious multi-file work, the larger Qwen3.5 tiers are probably a better fit than expecting the 9B model to behave like a cloud-grade coding assistant.