Local Qwen3.5 4B tops Cursor, Composer

The interesting signal here is not that a 4B model suddenly became frontier-grade everywhere; it is that agent wrappers still lose badly when they optimize for format compliance instead of actual correctness. For AI coding teams, this is a reminder that eval design and verification logic matter at least as much as raw model size.

• –The reasoning failure mode is concrete: the post shows Cursor Auto and Composer 1.5 missing the negative-floor edge case in a modular arithmetic sum, then drifting into inconsistent totals.
• –The frontend comparison matters because the author judged rendered React and Tailwind output, not just text answers, and says Qwen's page had better hierarchy, spacing, gradients, and interaction polish.
• –The setup is practical rather than lab-only: a 4-bit quantized local model on an RTX 3070 Mobile at roughly 55 tok/s, routed from LM Studio into Cursor with ngrok.
• –This is still an anecdotal benchmark from one user with one prompt bundle, so the real takeaway is to run your own evals before concluding that local small models broadly beat hosted coding agents.