The open-source model router that learns from every request, runs on your infrastructure, and puts every AI model at your fingertips — better than Sakana Fugu, zero vendor lock-in, zero monthly fees.
Three strategies — similarity, cost, and fallback — pick the best model for every prompt in under 20ms.
🔄
Feedback Loop
Learns from every request. Auto-retrains after every 10 requests. Gets smarter over time.
🔍
Fully Transparent
Every routing decision is visible. No black box. See which model was picked and why.
🏠
Self-Hosting
Run entirely on-premise with local models via Ollama. Air-gapped capable. No cloud dependency.
📊
Live Dashboard
Real-time web dashboard showing routing decisions, model usage, costs, and latency.
🔌
OpenAI-Compatible
Drop-in replacement for OpenAI API. Works with any OpenAI SDK. 100+ providers via LiteLLM.
Fugusashi vs Sakana AI Fugu
Two different bets on the future of AI routing.
Feature
Sakana Fugu
Fugusashi
Architecture
Trained 7B coordinator LLM
CMA-ES evolved weights + ensemble
Orchestration
TRINITY/Conductor (ICLR 2026)
Rule-based + GRPO learning
Self-Hosting
❌ Cloud-only
✓ Local-first
Cost
$5-30/M tokens
✓ Free
Transparency
❌ Black box
✓ Every decision visible + explained
Feedback Loop
❌ Static
✓ Learns from your traffic
Federated Learning
❌
✓ Collaborative routing
Model Pool
Sakana-controlled
✓ You control (100+ providers)
EU Availability
❌ Blocked (GDPR)
✓ Available everywhere
License
Proprietary
✓ MIT
Fugu bets on a trained coordinator with frontier models. Fugusashi bets on transparency, self-hosting, and control. Both are valid — pick based on what matters to you.
Quickstart
Up and running in 5 minutes.
# Install from PyPI pip install fugusashi
# Or run with Docker docker run -p 6060:6060 ghcr.io/eulogik/fugusashi:latest
# Or from source git clonehttps://github.com/eulogik/fugusashi.git cdfugusashi && pip install -e .
# Run fugusashi serve --config config.yaml
# Use curlhttp://localhost:6060/v1/chat/completions-H"Content-Type: application/json"-d'{"model":"auto","messages":[{"role":"user","content":"Say hello"}]}'
# Open dashboard openhttp://localhost:6060/dashboard
Architecture
Two tiers. One fast router. One deep orchestrator.
CMA-ES Coordinator
The learning brain. Evolves to pick the best model for every prompt.
📝
Prompt
→
🧠
Embed
→
⚡
Route
→
📊
Evaluate
→
🧬
Evolve
↺
Inspired by Sakana AI's TRINITY paper. Uses CMA-ES (Covariance Matrix Adaptation Evolution Strategy) to evolve routing weights. Runs locally in under 1ms per decision. Gets smarter from every request.
Feedback Loop
Learns from every request. Gets smarter over time.
🧠
Route
→
⚡
Execute
→
📊
Evaluate
→
📚
Learn
↺
Outcomes are recorded automatically. After every 10 requests, the similarity router retrains. Over time it learns which prompts work best with which models — without any manual intervention.
Benchmarks
80.0%
Held-Out Routing Accuracy
36.7%
Cost-Only Baseline
22ms
Median Routing Latency
100+
Supported Providers
Federated Routing
Collaborative routing without data sharing.
🏥
Hospital A
+
🚀
Startup B
+
🏢
Enterprise C
→
🧬
Federated Avg
→
🧠
Global Router
Each organization trains locally on their own data. Only weight updates (with differential privacy noise) are shared. The result is a router smarter than any single deployment.
Human-Interpretable Explanations
Every routing decision comes with a natural language explanation.
POST /v1/explain {"prompt": "Write a Python class for a binary tree"}
Decision: Route to gpt-oss-120b (confidence: 87%) Why: This prompt involves code generation. gpt-oss-120b is best suited for complex reasoning.
Alternatives considered: - lfm-2.5-1.2b (12%): better for fast responses - hermes-3-405b (1%): better for creative writing
Latency: 5.5ms | Strategy: cma-es
Paper
Research contributions: federated routing learning, human-interpretable routing, CMA-ES adaptation.
