OpenFugu: self-host a router that picks the best model per query
The closed Fugu orchestrator now has a free twin you can run locally: one endpoint over your own model pool, a router you train with CMA-ES or GRPO.
Copy markdown
One endpoint, many models
OpenFugu serves an OpenAI-compatible /v1/chat/completions that hides an internal pool — a ~0.6B Qwen3 backbone scores each worker and the top pick answers. Point any OpenAI SDK at it, no code changes, no vendor lock-in.
Train the router yourself
The picker is a bias-free linear head (~19.5K params) trained gradient-free with sep-CMA-ES, so it runs on a laptop; the Fugu-Ultra "Conductor" that emits workflow DAGs trains via GRPO on ToolScale. The repo claims +107% over the best single worker on its routing benchmark.
What it copies — and what the paid version costs
Sakana's own Fugu is API-only: $20–$200/mo or ~$5/$30 per Mtok on Ultra, offered via OpenRouter and Vercel AI Gateway, and blocked in the EU/EEA. OpenFugu is Apache-2.0 (weights under the Llama 3.2 license), so you get the same orchestration on your own keys.