Decima Guard puts its smaller prompt checker in your browser

Decima’s multilingual prompt checker now has a dedicated browser playground, including a smaller 140-million-parameter option. The October 19 addition lets visitors inspect a user message, retrieved document or tool result, model reply, or web input without first installing the Python package.

The project’s English description says inference runs locally through ONNX Runtime Web and WebAssembly. The selected int8 model downloads on the first decision and is then cached on the device: approximately 187 MB for Decima Guard Small or 370 MB for the 321M guard. The author says text stays on the device. That is the project’s documented architecture, not an AITrending network audit.

A smaller checkpoint, with a measurable tradeoff

The small preview was uploaded on October 18, a day before this browser addition. Its card reports 84.8% accuracy on the same 24,832-example public LLM-safety test used for the larger guard, versus 87.4% for the 321M version. On the prompt-injection subset, the figures are 89.6% and 91.5% respectively.

The author reports 50 milliseconds per short-message question on one CPU thread for the small model, compared with 111 milliseconds for the larger one. Those are project-reported CPU measurements, not browser latency promises. The live page describes a first browser check taking a few seconds and later checks about a second. Its browser runtime also uses float activations for int8 layers, so close-call probabilities can differ from native Python inference.

For applications, the documented route remains pip install decima-ai, with Router(model="amyrmahdy/decima-guard-small") selecting the smaller checkpoint. The playground exposes presets and probability outputs rather than granting or enforcing an agent’s execution permissions.

What this preview can and cannot establish

The browser page’s examples are author-written demonstrations, not held-out test cases. Both model sizes retain the preview’s limitations, including missed attacks and false alarms; the small card explicitly says its limitations are a little worse. The full-size card warns about Persian attack detection, misleading scores on ordinary web paths, and experimental source-code vulnerability checks.

The checkpoints declare Apache-2.0, while the guard documentation discloses some training datasets without published licenses. Users who depend on fully resolved training-source rights should read that qualification.

This is a concrete extension of the earlier CPU preview: a smaller checkpoint and an accessible local browser workflow. It is not a newly created Space, a guarantee of security, or evidence that AITrending reproduced the author’s benchmark results.