Qwen3.5-9B stalls in agent mode

Qwen3.5 is supposed to be agent-friendly, so a silent stop usually means the wrapper or runtime is the weak link. This looks more like a context-budget or tool-format mismatch than a pure hardware limit.

• –Qwen’s docs say Qwen3.5 thinks by default and needs explicit non-thinking config for direct responses, which some wrappers handle poorly.
• –Qwen3.5 has a 262K native context, so a runtime with a tiny default window can choke the conversation long before the model reaches its real limit.
• –The same stall on 4B makes raw RAM a weaker explanation than context length, stop-sequence handling, or parser mismatch.
• –Heavy agent harnesses like OpenCode and Claude Code send large prompts and tool schemas every turn, so smaller local models can appear “done” when the conversation budget is already gone.
• –For agentic work, Qwen points developers toward current serving frameworks and Qwen-Agent rather than generic wrappers that may drop reasoning content.