little-coder lifts Qwen into top 10

This is less a clean model victory than a scaffold warning shot: local models may look weaker partly because they are tested inside agents tuned for frontier-model behavior.

• –The result puts a 35B-total, 3B-active MoE model in the public top-10 band on Aider Polyglot, which is unusually strong for an offline local setup.
• –The biggest gain came from first-attempt solves, suggesting Qwen3.6-35B-A3B is doing more than benefiting from retry mechanics.
• –little-coder’s small-model guardrails, including tool-use constraints, workspace discovery, and reasoning-budget control, make the harness part of the benchmark result.
• –The methodology is still self-reported and benchmark-specific, so Terminal Bench and GAIA follow-ups will matter before generalizing the claim.
• –For developers running local agents, this points toward optimizing scaffolds, prompts, and tool loops before assuming only larger cloud models can compete.