Mahoraga says Qwen3 4B tops cloud agents

The interesting part is not just that a local model wins on one benchmark slice, but that Mahoraga turns routing itself into a learnable system instead of a hand-tuned ruleset. That makes this more compelling as infrastructure than as a one-off model leaderboard post.

• –The strongest claim is narrow but useful: Qwen3 4B looks best for code/refactor, while other agents still win on research or planning buckets.
• –The LinUCB setup is the real product story: it learns per-bucket routing over time, which is exactly the kind of adaptation static “best model” rules miss.
• –The benchmark design is practical: no LLM-as-judge, no paid eval loop, and a local hardware target that matches the “I have a laptop, not a datacenter” constraint.
• –The weak spot is the scorer itself: security came out flat across agents, so the system still needs better task-specific evaluation before anyone should trust it broadly.
• –This will matter most to developers who care about cost control, offline workflows, or shuttling tasks between cheap local models and premium cloud models automatically.