GLiNER2 Author Pushes Back on Jev Hype, Highlights Open GLiGuard Guardrails Model

philipvollet · x · 2026-09-20

The GLiNER2 author responds to attention around Jev, arguing its schema-conditioned multi-task classification isn't new: GLiNER2 (EMNLP 2025, Apache-2.0) has done user-defined label spaces with confidence scores in one forward pass since 2025. They also highlight GLiGuard (fastino/gliguard-LLMGuardrails-300M, 0.3B), a CPU-first guardrails model that treats safety moderation as schema-conditioned classification—covering prompt safety, toxicity, and jailbreak detection in a single encoder pass—claiming accuracy competitive with models 23-90x larger while running up to 16x faster with 17x lower latency.

Original post →

More from Models

Models channel →