Tasks and inputs
Shieldstral 1.0 is listed for Content Moderation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Oct 7, 2026
Shieldstral 1.0 is a compact open-weight moderation model that classifies text and image inputs against natural-language policy questions.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Shieldstral 1.0 is a compact open-weight moderation model that classifies text and image inputs against natural-language policy questions.
The publisher describes prompt moderation, response moderation, prompt-response pair checks and refusal detection. The output is a yes-or-no classification. It belongs in an evaluation or filtering step; it should not be presented as a general model that writes a replacement answer.
A vague policy question makes the result difficult to interpret. Write the category and decision rule your product needs, then assemble positive cases, negative cases and borderline examples. Distinguish quoted material, educational discussion and the action a user is requesting. Track false positives alongside missed cases.
The documented checkpoint has 3.8B parameters and a 32,000-token context allowance, with Apache 2.0 licensing. It is in public preview. Check the supported serving configuration and the format expected for combined text-and-image inputs before adding it to an existing moderation pipeline.
Keep an auditable record of the policy version, input type and returned classification. Changes to the policy question or upstream content formatting can change what a classification means. Validate each change on a held-out set instead of assuming an old result applies.
For model selection, compare moderation classifiers within the content moderation category. General reasoning models and dedicated policy classifiers serve different roles, so their output quality needs different acceptance criteria.
Publisher source 1 · Publisher source 2 Deployment limits and commercial terms belong to the selected checkpoint or API route.
Compare the model on the work you need to complete.
Shieldstral 1.0 is listed for Content Moderation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 32K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
OpenAI
GPT OSS Safeguard 120B from OpenAI: Safety model for policy screening, moderation, and risk-aware routing workflows.
OpenAI
GPT OSS Safeguard 20B from OpenAI: Safety model for policy screening, moderation, and risk-aware routing workflows.
Meta
Llama-Guard-3-8B from Meta: Llama 3.1-based safety classifier for moderating prompts and model responses.