NVIDIA: Nemotron 3.5 Content Safety (free)

Nemotron 3.5 Content Safety is NVIDIA’s 4B-parameter guardrail model, fine-tuned from Gemma-3-4B. Its documented role is prompt and response moderation, content classification and safety-policy workflows. NVIDIA describes outputs that assess a user prompt and response as safe or unsafe, add safety-category labels and optionally include a reasoning trace. The assessment is model output for the application to evaluate and apply. Each message has a role of system, developer, user or assistant and plain string content. An optional name is a string; keep each message to these fields. Send stream:false, or omit stream.

Use the exact model ID nvidia/nemotron-3.5-content-safety:free with POST /v1/chat/completions and a nonempty messages array. Put the material to assess in text content and distinguish the user prompt from the model response when both are being reviewed. A practical moderation step can label those two passages explicitly and ask the model to identify the relevant categories while keeping its assessment tied to the supplied content.

This Ailoot edition accepts text and returns text. For a review queue, prepare the actual post, support-message draft or generated answer that needs assessment, together with the surrounding context that changes its meaning. Keep application decisions such as routing a flagged item for human review in your own workflow. Media inputs, tool-calling controls and explicit cache controls are outside this edition’s request profile.

The model page lists a 128,000-token context window and up to 8,192 completion tokens. Ailoot also requires the entire UTF-8 JSON request body, including instructions and field names, to fit within 128,000 bytes. For an explicit output cap, set exactly one of max_tokens or max_completion_tokens to a positive integer from 1 through 8,192. If neither is supplied, Ailoot forwards a 4,096-token output cap. Account for the complete conversation and the assessment you request. The model description reports coverage of 12 languages; preserve the original wording when the evaluation depends on the source language. If your application expects particular labels, parse and validate the returned text against your chosen handling rules. A requested JSON-looking answer is not a guarantee of schema-enforced output.

Use stream:false for a complete JSON conversation response. Read choices[].message.content, preserve every returned choice and any supplied reasoning fields, and retain finish_reason and usage. Store the assessment with the exact prompt and response it evaluates so the decision can be reviewed later. An assessment of one text applies to that supplied text and context; evaluate a materially changed response as its own item.

Prepare material that contains no confidential information, as required by this free edition’s documented usage terms. If the request ends without a complete assessment, retain the original request identifier and returned data while checking the existing attempt. This card documents a direct conversation result and supplies no separate asynchronous job or cancellation workflow.