Fastino releases one model for guardrails and PII detection
Fastino has released GLiNER2-Guardrails-PII-Multi, an open-source 0.3B-parameter model that does safety moderation and PII extraction in one forward pass, so agent pipelines no longer need two separate filter models.
Fastino announced GLiNER2-Guardrails-PII-Multi, a multilingual, multi-task small language model with 0.3 billion parameters. The model combines content-safety moderation and privacy filtering, two jobs that normally run on separate models. According to the company, it matches the performance of each of its predecessors while cutting inference time and memory use in half.
The model merges two earlier Fastino releases. GLiGuard is a zero-shot guardrail model with a safety policy the company compares to QwenGuard's. GLiNER2-PII is a privacy filter that detects 42 types of personally identifiable information. Fastino says the two models passed 52,000 downloads on Hugging Face within six weeks of release.
On the safety side, the model runs four tasks at once. It gives prompts and responses an overall safe/unsafe label, detects 11 jailbreak strategies in prompts, flags 14 harm categories in both prompts and responses, and detects refusals in responses. On the privacy side, it extracts 42 PII types in seven categories, including government IDs, banking and payment details, and secrets such as API keys and access tokens.
The model is built on the GLiNER2 encoder architecture. That architecture places the task definitions in front of the input text and scores labels against text spans using dot-product similarity, which lets the model handle classification and extraction in a single pass. Fastino says it trained the model on enhanced versions of the original datasets and generated synthetic data with its Fine-Tuning Agent API, because real PII training data is scarce.
The available excerpt does not include benchmark figures, a list of supported languages, or license details.