GLiGuard by Fastino Labs: a safety model 16x faster than larger competitors

Fastino Labs released GLiGuard on May 13, 2026 — a 300-million-parameter safety moderation model that inspects a prompt across four protection directions in a single pass and works an order of magnitude faster than comparable alternatives.

What GLiGuard checks in one pass

GLiGuard solves four moderation tasks in a single forward pass: it evaluates the safety of the request itself, detects signs of jailbreak attempts, determines the category of potential harm, and predicts whether the model will refuse to respond. Previously, such checks were typically distributed across separate models or multiple sequential passes of one model, which increased production latency.

• Model size — 300 million parameters
• Four moderation tasks in one forward pass: prompt safety, jailbreak strategy detection, harm category classification, refusal detection
• Weights published on Hugging Face under Apache 2.0 license
• Accuracy verified on nine safety benchmarks
• Comparison conducted with models 23–90 times larger than GLiGuard

Why encoder architecture delivered speed

Fastino Labs built GLiGuard on an encoder architecture rather than decoder-only, which is used by most existing guardrail models. According to MarkTechPost, this architectural choice gave the model up to 16 times higher throughput and 16.6 times lower latency compared to current top-tier moderation models.

"GLiGuard achieves up to 16-fold increase in throughput and 16.6 times

lower latency compared to state-of-the-art models, while matching or exceeding the accuracy of models 23–90 times larger," according to MarkTechPost.

How accurate is the model at this size

GLiGuard at 300 million parameters matches the accuracy of models 23–90 times larger, and on some of nine tested safety benchmarks exceeds them. Fastino Labs released the model weights in open access on Hugging Face under Apache 2.0 license — GLiGuard can be used and modified for free, including in commercial products.

What this changes for production

Low latency and compact size make GLiGuard a candidate for real-time checks — where each additional model in the pipeline adds noticeable response delay. Since weights are distributed under Apache 2.0 license, companies can deploy GLiGuard internally and avoid per-request licensing fees, as often occurs with proprietary guardrail APIs.

What this means

GLiGuard demonstrates that prompt moderation doesn't require a giant model: a compact encoder-based architecture can deliver top-tier guardrail accuracy with significantly lower latency and computational cost, making it easier to integrate such protection into production pipelines that operate in real time.

Frequently asked questions

Where can I download GLiGuard weights?

The model is available on Hugging Face under Apache 2.0 license — it can be downloaded and used for free, including in commercial projects.

How does GLiGuard differ from other guardrail models?

Most guardrail models are built as decoder-only, while GLiGuard uses an encoder architecture — this, according to MarkTechPost, provided a 16-fold boost in throughput and 16.6-fold reduction in latency.

Which tasks is GLiGuard suited for?

Based on its stated metrics — low latency and high throughput while maintaining accuracy — the model is designed for scenarios where prompt filtering must be performed at scale and speed: chatbots, agent pipelines, and other production systems with high traffic.