apex-flash-1
Cantina released apex-flash-1, an MIT-licensed open-weights security model fine-tuned from GLM-5.3-Flash that solved 40 of 60 held-out vulnerability tasks.
Cantina fine-tuned GLM-5.3-Flash with reinforcement learning on 150 tasks built from 50 real vulnerability cases, each in three variants with different amounts of information. On 60 held-out tasks it passed 66.7%, against 60% for the untuned base model and 71.7% for Claude Opus 5 High. Cantina also released an abliterated variant with reduced refusals.
- Date
- Thursday 1 October 2026
- Lab
- Cantina
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Held-out tasks solved | 40 of 60 (66.7%) Claude Opus 5 High 43 of 60, base GLM-5.3-Flash 36 of 60 | company |
| Evaluation run cost | $2.38 reported $74.68 for Claude Opus 5 | company |
| Parameters | 321B | company |
Evaluation is internal, single runs on Cantina-built environments. Search results mention Yeta Labs as a build partner and an Oct 1 announcement date.
Sources
- runtimewire.com/article/cantina-apex-flash-open-weights-security-model
- marktechpost.com/2026/10/04/can-an-open-model-do-security-research-cantinas-apex-flash-1-s
This record was checked against its sources on 8 October 2026. How we check