What is GLiGuard?

On May 12, 2026, Fastino Labs released GLiGuard, a 300 million parameter open-source safety moderation model. Unlike decoder-only guardrail models that generate verdicts autoregressively, GLiGuard reframes safety moderation as a text classification problem, eliminating the sequential latency bottleneck [citation:4].

Key Features

  • Four Tasks in One Pass: Safety classification, jailbreak detection, harm categorization, and refusal detection
  • Encoder-Based Architecture: Eliminates sequential latency of decoder-only models
  • Open Source: Apache 2.0 license on Hugging Face
  • Single GPU Deployment: Runs efficiently without heavy infrastructure

Benchmark Performance

  • Prompt Classification: 87.7 average F1 (within 1.7 points of best model)
  • Response Classification: 82.7 average F1 (second highest overall)
  • Throughput: Up to 16.2× higher (133 vs 8.2 samples/s at batch size 4)
  • Latency: 16.6× lower — 26ms vs 426ms at sequence length 64
  • Outperforms: LlamaGuard4-12B, ShieldGemma-27B, NemoGuard-8B despite being 23-90× smaller

Pricing

Free and open source under Apache 2.0 license. Model weights available on Hugging Face at fastino/gliguard-LLMGuardrails-300M.

Pros

  • Exceptionally fast (26ms inference time)
  • Matches or beats models 90x larger
  • Truly open source with permissive license
  • Single GPU deployment
  • Multi-task in single forward pass

Cons

  • New model with limited community adoption yet
  • Encoder architecture may be less familiar to some developers
  • English-focused (multilingual capabilities unclear)
  • No commercial support options

Who Should Use It?

Perfect for: LLM application developers needing low-latency safety filtering, real-time chat moderation, and efficient guardrail deployment.

Verdict

GLiGuard is a remarkable achievement in efficient AI safety. The combination of speed, accuracy, and tiny size makes it the best option for production guardrails.

Rating: 4.6/5 - The new standard for safety moderation.