Fastino recounts GLiNER's path to 45M downloads

Fastino Labs has published an origin story for GLiNER, the open extraction model family whose latest release, GLiNER2.5, we covered earlier. The post adds history and adoption figures rather than a new model.

GLiNER began as a PhD project by Urchade Zaratiana at Université Sorbonne Paris Nord. He is now a founding researcher at Fastino. The idea was that entity extraction did not need a generative LLM. "All the information is already present in the input text," Zaratiana said. "You don't need to generate, you just need to find a way to extract these entities." The model places entity-type embeddings and span embeddings in a shared latent space, which lets a small encoder run on a CPU.

According to Fastino, GLiNER has passed 45 million downloads on Hugging Face in three years. Users include NGOs working on PII detection and humanitarian projects, UNESCO for guardrails, and university researchers.

The post traces three generations of the model:

• The 2023 original handled only named entity recognition. Spans were capped at 12 words and inputs at 512 tokens.
• GLiNER2 (2025) added classification, JSON extraction and relation extraction in one forward pass. Derivatives include GLiGuard for safety moderation, GLiNER2-PII, which covers 42 PII types, and a 0.3B model that combines the two.
• GLiNER2.5 predicts where spans start and end directly, removing the length limit. It raises the input window to 4,096 words and adds constrained classification to keep labels consistent across tasks. It comes in 74M, 0.2B and 0.3B multilingual variants under Apache 2.0.

The post does not include new benchmark results.