CVE-2026-124994

Invalid model metadata could trigger an out-of-bounds GPU read

Turboderp · ExLlamaV3 NGramEmbedding model loader

First documented disclosure · Review checkpoint

Summary

The model loader failed to check an index against the available embedding rows. Model processing could read past the intended memory region.

Severity and exploitation evidence

Medium · 5.3/10 · CVSS 4.0. Severity source.

Exploitation not established: The CNA reports that a demonstration is public. It does not report malicious exploitation.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The model loader failed to check an index against the available embedding rows.

Conditions required

An affected runtime loaded a model package containing inconsistent metadata.

Potential impact

Model processing could read past the intended memory region.

Affected versions and configurations

Through 1.6.3 according to the CNA.

Fix and mitigation status

1.6.4.

Practical next steps

Update ExLlamaV3 and use trusted model packages.

Operating lesson

Check the authority granted to the affected component against the operation a user expects.

Evidence and limits

The public project issue dates to October 17. The CNA published on November 9 and identifies 1.6.4 as corrected. The public disclosure was on 2026-10-17. The CVE record was published on 2026-11-09.

Sources