Model Precision Metadata API Guide

Model precision overview

Read the selected provider's metadata

Precision belongs to an endpoint, not a model name. GET /v1/models/{author}/{slug}/endpoints returns quantization and trustedrouter.precision for each endpoint. The catalog also exposes these at trustedrouter.endpoints[].quantization and trustedrouter.endpoints[].precision.

import httpx

model = "openai/gpt-oss-120b"
response = httpx.get(
    f"https://trustedrouter.com/v1/models/{model}/endpoints", timeout=20
)
response.raise_for_status()
for endpoint in response.json()["data"]:
    precision = endpoint.get("trustedrouter", {}).get("precision")
    print(endpoint["provider"], endpoint.get("quantization"))
    if precision:
        print(precision["reviewed_on"], precision["model_revision"])
        print(precision["sources"])

These are public metadata endpoints and require no API key. Use model pages to inspect the same evidence visually.

Field meanings

• quantization: the primary reviewed weight format. Unknown values are null.
• weight_formats, label, notes: mixed formats and higher-precision exclusions.
• kv_cache_dtype: the cache representation, separate from model weights. An FP8 KV cache does not establish FP8 weights.
• model_repository, model_revision: the source model and reviewed revision.
• reviewed_on, sources: the review date and links to serving code and model configuration.
• evidence_type: published_serving_config.
• runtime_verified: false. This is not per-request proof of the weights loaded by the upstream provider.