Saina Helm 2 Routing 0.8B

A text-only routing specialization of Saina Helm. It chooses among candidate models from a request, capability descriptions and estimated prices, returning a probability for each option instead of generating text.

This repository contains standalone merged BF16 weights, the decision head and tokenizer. No separate adapter is required.

Hosted API

Select saina-helm-2-routing-0.8b at https://api.saina.run/v1/ask. The model registry is available at https://api.saina.run/v1/models.

import os
import requests

response = requests.post(
    "https://api.saina.run/v1/ask",
    headers={"Authorization": "Bearer " + os.environ["SAINA_API_KEY"]},
    json={
        "model": "saina-helm-2-routing-0.8b",
        "state": "Write a short thank-you email after a meeting.",
        "questions": {
            "route": {
                "type": "single_choice",
                "question": "Select the cheapest option likely to achieve quality at least 0.90 on this request. If none can meet 0.90, prioritize highest quality, then lower cost. Capability scores are averages on independent tasks; prices are estimates.",
                "options": {
                    "economy": "Text model. Suitable for short everyday writing. Estimated request cost: $0.0001.",
                    "advanced": "Text model. Suitable for complex reasoning and demanding writing. Estimated request cost: $0.001."
                }
            }
        }
    },
    timeout=60,
)
response.raise_for_status()
print(response.json())

Option identifiers are application-defined. Put capability information and estimated request costs in the descriptions; model names are not required. Replace the illustrative descriptions and prices with information about your actual candidates.

Local use

Install saina[local]>=0.4.0 and load this checkpoint with an immutable Hub revision. See the Saina SDK for local inference and the API documentation for request formats.

This is an experimental release. Measure routing quality and total cost on your workload before relying on it. Returned probabilities describe the routing decision; they are not a guarantee of downstream answer quality.