I Built a Tiny Local Model That Outperforms Opus! (by using Orchestration)

PandaCode: https://github.com/aicodeking/pandacode PandaOrchestrator: https://github.com/aicodeking/PandaOrchestrator In this video, I'll show you PandaOrchestrator and Panda Router, a tiny local model I built to decide which model each worker in an agent team should use. I'll go through how it works, how I trained it on a budget with Ettin, Modal and an attempt-and-grade setup, and the numbers on 150 held-out tasks, including the escalation ladder that gets it close to perfect routing. Then I'll show all the new updates inside PandaCode: the router built into the app, the Team menu with Cheap, Balanced and Best teams, failover, model health, the model tier list, sub-workers, team messaging and the leaner architect. Finally, I'll run a Balanced team on a real task and give you my honest take on where it still falls short. -- Key Takeaways: 🧠 Panda Router is a tiny decision model, not a chatbot: it profiles a task and plain code scores every model you can run on success, cost and speed. 🗂️ Models are just JSON cards, so when a new model comes out you add a card and never retrain the router. 💸 The 32M student model is about 31 MB in int8 ONNX, routes in 10 to 25 ms on a CPU, and its Modal training run cost $1.67. 🔍 DeepSeek V4.1 Flash labelled about 8,000 tasks for $2.77, and real attempts graded by a strict judge gave the router true outcome data. 📊 On 150 held-out tasks, one pick solved about 77 out of 100, two attempts on the escalation ladder reached about 91, and perfect routing is 94. 💻 The router now ships inside PandaCode in Rust with ONNX Runtime, no Python or server needed, and picks from every installed model. 👥 The Team button gives you No team, Cheap, Balanced and Best team, and failed workers only ever step up to a stronger model. ⚡ Failover moves a broken worker to the same model on another harness, and model health pauses failing models with an Allow now button. 🎯 The tier list lets you drag models into S to D, and your ranking steers the router's success estimate 70/30. 👍 Verdict: it's not perfect, but it stops wasting your best models on easy stuff and weak models on hard stuff, and it's all open source.