Tiny Browser-Run Security Model Gets a Fix for Four-Bit Output

A small model for a narrow job

Devseis published a v0.6 build of its 0.5B Endpoint Auditor model on October 13, with a concrete fix for running it in a four-bit browser format. The project’s English model card says v0.5 performed acceptably at full precision but produced poor output after quantization. For v0.6, the author retrained with the rounding behavior of the target build inside the training loop. The model is meant to draft risk explanations, remediation steps and summaries from endpoint evidence while running locally through WebLLM.

Verdicts still come from code

Devseis says application code collects the evidence and determines each compliance verdict or vulnerability match; the model writes the explanation and its output is checked against that evidence before use. That division matters because a generated sentence is not itself an audit finding. The card reports better grounded results for the four-bit build on the author’s held-out prompts, but those small tests are not independent proof of broad reliability or formal ISO 27001, GDPR or EU AI Act compliance. The new contribution is an available specialized community model and a documented quantization repair, not a claim that it can certify an organization by itself.