Article
Aikido cuts one-third of GLM-5.3's experts to create a local security model
Beating AI News: Belgian cybersecurity company Aikido released its first open-source security model, Altar-1. It turns Z.ai's GLM-5.3 into a compressed version for security scenarios. Enterprises can deploy it on their own servers or even in completely offline environments for code auditing and penetration testing, without handing source code or unpatched vulnerabilities to external models.
Aikido first converted GLM-5.3 into an INT4-quantized version, then used Cerebras's REAP method to prune 256 routing experts down by 88, equivalent to cutting 34.4%. The model size was reduced from 1.51TB in full precision to 328GB, which is 32.8% smaller than the 488GB quantized version. The official recommendation is to deploy it with 4 H200s.
Performance was largely preserved after compression. In Aikido's self-built test of 32 CVEs, Altar-1 achieved an average single-pass recall of 60.4%, compared with 61.5% for quantized GLM-5.3. Across three consecutive tests, both found 23 vulnerabilities at least once, while full-precision GLM-5.3 found 25.
Strictly speaking, Altar-1 is currently more like GLM-5.3 pruned for targeted security scenarios. Aikido explicitly stated that this step did not train any new capabilities; fine-tuning for actual security workflows, tool use, and long-horizon reasoning is still to come.
CAISI, under NIST, previously rated the base GLM-5.3 as the strongest open-source model in terms of cybersecurity capabilities, although its overall capabilities still lag behind current leading U.S. models by approximately four months.
Original link: https://m.theblockbeats.info/flash/368428