Fastino Labs' GLiGuard: A Safety Model 16x Faster Than Much Larger Rivals

On May 13, 2026, Fastino Labs released GLiGuard as open source: a safety moderation system with 300 million parameters that checks a prompt on four protection dimensions in a single pass and runs an order of magnitude faster than comparable tools.

What GLiGuard checks in one pass

In a single forward pass, GLiGuard handles four moderation tasks: it assesses the safety of the request itself, looks for signs of a jailbreak attempt, determines the category of potential harm and predicts whether the model will refuse to answer. Previously, such checks were usually split across separate models or several sequential passes of one model, which increased latency in production.

• 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 the Apache 2.0 license
• Accuracy tested on nine safety benchmarks
• Compared against models 23 to 90 times larger than GLiGuard

Why the encoder architecture delivered the speed

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

"GLiGuard achieves up to 16x higher throughput and 16.6x lower latency than current state-of-the-art models, while matching or exceeding the accuracy of models 23 to 90 times larger,"

MarkTechPost writes.

How accurate is a model this small

At 300 million parameters, GLiGuard does not fall behind models 23 to 90 times larger in accuracy, and on some of the nine safety benchmarks tested it outperforms them. Fastino Labs published the weights on Hugging Face under the Apache 2.0 license, so GLiGuard can be used and modified for free, including in commercial products.

What this changes for production

Low latency and a compact size make GLiGuard a candidate for real-time checks, where every extra model in the chain adds noticeable response delay. Because the weights are distributed under Apache 2.0, companies can deploy GLiGuard themselves and avoid paying per-request licensing fees, as is often the case with proprietary guardrail APIs.

What this means

GLiGuard shows that prompt moderation does not require a giant model: a compact encoder-based architecture can deliver the accuracy of top guardrail systems at a fraction of the latency and compute cost, which makes this kind of protection easier to build into real-time production pipelines.

Frequently asked questions

Where can I download the GLiGuard weights?

The model is available on Hugging Face under the Apache 2.0 license. You can download and use it for free, including in commercial projects.

How does GLiGuard differ from other guardrail models?

Most guardrail models are decoder-only, while GLiGuard uses an encoder architecture. According to MarkTechPost, this is what delivered the 16x gain in throughput and the 16.6x gain in latency.

What tasks is GLiGuard suited for?

Based on the stated metrics, low latency and high throughput with accuracy preserved, the model is designed for scenarios where prompts must be filtered quickly and at scale: chatbots, agentic pipelines and other high-traffic production systems.