eulogik/fugusashi-router

Fugusashi Router — CMA-ES Evolved Model Routing Weights

Like Sakana Fugu. But Free. And Yours.

---

What is This?

CMA-ES evolved routing weights for the Fugusashi intelligent model router — an open-source alternative to Sakana AI's Fugu model router.

A 385-dimensional weight vector (384 embedding weights + 1 bias) that maps prompt embeddings to model selection, learned via Covariance Matrix Adaptation Evolution Strategy (CMA-ES) — the same evolutionary optimization approach used in Sakana AI's TRINITY paper.

This is the evolved router brain — the weights that tell the Fugusashi router which model to pick for each prompt.

Quick Start

import numpy as np
import json
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download

# Download weights from HuggingFace
weights_path = hf_hub_download(repo_id="eulogik/fugusashi-router", filename="cmaes_weights.json")
with open(weights_path) as f:
    data = json.load(f)

weights = np.array(data["mean"])
bias = weights[-1]

# Embed prompt and route
embedder = SentenceTransformer("all-MiniLM-L6-v2")
embedding = embedder.encode(["Write a Python function to sort a list"], normalize_embeddings=True)[0]

logits = embedding @ weights[:384] + bias
models = ["gpt-oss-120b", "nemotron-3-ultra", "nemotron-3-super", "hermes-3-405b", "lfm-2.5-1.2b"]
chosen = models[np.argmax(logits)]
print(f"Route to: {chosen}")

Or Install the Full Router

pip install fugusashi

from fugusashi.coordinator import CMAESRouter
from fugusashi.config import load_config

config = load_config("config.yaml")
router = CMAESRouter(config=config)
result = router.route("Write a Python sorting function")
print(f"Model: {result.model}, Latency: {result.latency_ms:.1f}ms")

Training Details

Field	Value
Algorithm	CMA-ES (Covariance Matrix Adaptation Evolution Strategy)
Dimensions	385 (384 embedding weights + 1 bias)
Population	16
Generations	30
Training tasks	20 preference samples
Best fitness	~0.24
Embedding model	all-MiniLM-L6-v2
Mean latency	4.0ms
P95 latency	4.4ms
P99 latency	7.9ms

Performance

• 4ms mean routing latency — 25x faster than LLM-based routing
• 85% accuracy with training data, 70% cold-start
• CMA-ES evolution — weights improve continuously from outcomes
• Federated learning — privacy-preserving collaborative routing
• Natural language explanations — every decision explained

Model Pool

Model	Role	Provider
gpt-oss-120b	Heavy reasoning	OpenRouter (free)
nemotron-3-ultra	General	OpenRouter (free)
nemotron-3-super	Light general	OpenRouter (free)
hermes-3-405b	Code/creative	OpenRouter (free)
lfm-2.5-1.2b	Ultra-fast	OpenRouter (free)

How It Works

Prompt → Embed (384d) → Dot Product (weights[:384]) + bias → argmax → Model

The 385-dimensional weight vector is evolved via CMA-ES across 30 generations with a population of 16. Each generation evaluates routing accuracy on a training set of preference samples.

Project Links

Resource	Link
Source Code	github.com/eulogik/fugusashi
PyPI Package	pypi.org/project/fugusashi
Live Demo	huggingface.co/spaces/eulogik/fugusashi
Dataset	huggingface.co/datasets/eulogik/fugusashi-preferences
Website	eulogik.github.io/fugusashi
eulogik	eulogik.com

Citation

@software{fugusashi2026,
  title={Fugusashi: Open-Source Intelligent Model Router},
  author={{eulogik}},
  year={2026},
  url={https://github.com/eulogik/fugusashi},
  doi={10.5281/zenodo.XXXXXXX}
}

License

MIT — use it however you want.

---

Built with ❤️ by eulogik

Fugu Sashi. Served Free.