Automatic model choice (silicon/auto)

The names

Put one of these in model on a chat surface (chat completions, Responses, Claude Messages, the Ollama chat and generate routes). Names are case-insensitive.

curl -si https://api.siliconnetwork.io/v1/chat/completions \
  -H "Authorization: Bearer $SILICON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "silicon/auto", "messages": [{"role": "user", "content": "Name three prime numbers."}]}'

HTTP/2 200
silicon-model: mlx-community/Meta-Llama-3.1-8B-Instruct-4bit
silicon-routed-from: silicon/auto
...

{"model": "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit", "choices": [...], "usage": {...}}

The model in that answer is an example: which one serves depends on what is loaded on the network at that moment.

How the model is chosen

Every request is matched again, against what is true when it arrives:

1. Being served

Only a model that is loaded on the network right now counts: on the platform's engine, or on an online operator node that serves it. A key pinned to regions, operators or attested nodes counts only what its pins allow. The choice is made from the table below and nothing else, so a model that is not in it is never picked, however cheap.

2. The request's needs

The prompt estimate is deliberately high (about one token per three bytes of text, plus allowances for images and the chat template), so a model is skipped rather than sent a prompt it may not hold. Billing always uses the engine's actual token counts.

3. Price

Among the models that qualify, the price is each model's current per-token rate (the same rate it bills at, see the pricing page) applied to this request's estimated prompt and output tokens. The tier then orders them as above.

The quality rank is a rough editorial order of these models, not a benchmark score. Active is the parameters used per token, which differs from the total for a mixture of experts; it is what the fast tier ranks on. The two smallest Llama models are not offered for tool calls.

What the response says

The answer names the model that served it, never silicon/auto:

You are billed for the model that served, at its own price, exactly as if you had named it. The usage in your dashboard and the usage export name that model.

Fallback

The ranked list is the request's fallback chain: when the first pick cannot serve before any output reaches you (not loaded after all, no capacity, an upstream error or timeout), the next one is tried, up to 5 models in all. Only the attempt that answers is billed. A models array still works after it, for a fallback of your own:

from openai import OpenAI

client = OpenAI(base_url="https://api.siliconnetwork.io/v1", api_key="sn_live_...")
out = client.chat.completions.create(
    model="silicon/auto:best",
    messages=[{"role": "user", "content": "Summarise RFC 9110 in three lines."}],
    extra_body={"models": ["llama-3.1-8b"]},  # tried after the automatic picks
)
print(out.model)  # the engine id that answered

With quality (economy, standard, max) each model is considered in that quantization, and only if that variant is being served. See Models, aliases and fallbacks.

Keys scoped to models

A key whose scope lists models is only ever served one of them: silicon/auto picks among the models the key may call, by catalog alias or full id. A scope that lists silicon/auto itself allows whatever it picks from the table. A key's region, operator and attestation pins apply as they always do.

Errors

In GET /v1/models

While at least one model in the table is being served, GET /v1/models lists silicon/auto and its three tiers, each with a description, its tier, the models it would try for a request with no particular needs (resolves_to, best first) and the one it would use now (serving). A request with tools, images, JSON output or a long prompt may be served by another model.

{
  "id": "silicon/auto",
  "object": "model",
  "owned_by": "silicon-network",
  "description": "Automatic model choice, balanced tier: ...",
  "tier": "balanced",
  "resolves_to": ["mlx-community/Meta-Llama-3.1-8B-Instruct-4bit", "..."],
  "serving": "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
}