Parameters: 117B (5.1B active)Quantization: MXFP4 (native MXFP4 MoE weights, as released by OpenAI)Context: 131K tokensStrengths: Safety reasoning, bring-your-own-policy flexibility, full access to reasoning chains for debugging, configurable reasoning effort levelsStructured Outputs: Structured response formatting supportBest for: Content moderation, policy enforcement, LLM guardrails, and Trust & Safety labeling workflowsModel weights: openai/gpt-oss-safeguard-120bConfiguration repo: tinfoilsh/confidential-gpt-oss-safeguard-120b
Safety Model: Classifies text content based on custom safety policies you provide.
Parameters: 1.5B (50M active)Strengths: Token-level PII span detection with typed labels (person, email, phone, address, date, URL, account number, secret), CPU inference without generative promptingBest for: Detecting and redacting personal information in text, including outgoing web search queriesModel weights: openai/privacy-filterAnnouncement: Introducing OpenAI Privacy FilterConfiguration repo: tinfoilsh/confidential-pii-cpu
Use the authenticated privacy-filter API, or enable
pii_check_options in web search. The raw endpoint returns the upstream model’s spans and placeholder-redacted text; web search separately removes selected spans from outgoing queries. Detection is not a guarantee that all PII will be recognized.