OpenClaw qwen3.5:9b Echoes Claude Code

The real story is not “9B beats Claude,” it’s that agent orchestration can compensate for a lot of model weakness when the system is ruthless about transitions, compression, and tool discipline.

• –The strongest claim here is the hard cutoff: once tools are removed after a fixed number of exploration steps, the model stops looping and actually ships output.
• –qwen3.5:9b’s native tool_call structure is the key enabler; the post suggests the qwen2.5 line is much less reliable because it emits JSON in content instead of structured tool calls.
• –The prompt-compression and deferred-loading ideas are the kind of unglamorous infrastructure work that makes local agents feel fast enough to use.
• –The results are interesting, but they are still self-benchmarked and highly setup-dependent, so they read more like a strong systems writeup than a general model ranking.
• –OpenClaw’s Product Hunt presence frames it as a local, automation-first agent with real OS access, which makes the security and control tradeoffs in the post especially relevant.