Aikido prunes one-third of the experts from GLM-5.3 to build a local security model.

Beating AI Express News: Belgian cybersecurity firm Aikido has launched its first open-source security model, Altar-1. The model compresses Z.ai’s GLM-5.3 into a version optimized for security use cases. Enterprises can deploy it on their own servers, even in fully offline environments, for code audits and penetration testing—eliminating the need to share source code or unpatched vulnerabilities with external third-party models. Aikido first converted GLM-5.3 to an INT4 quantized version, then applied Cerebras’ REAP method to prune 88 of its 256 routing experts, a 34.4% cut. The model’s size shrank from 1.51TB (full precision) to 328GB, 32.8% smaller than the quantized GLM-5.3’s 488GB footprint. The official recommends deploying Altar-1 on 4 H200 GPUs, with performance largely preserved post-compression. In Aikido’s internal 32-CVE test, Altar-1 posted an average single-round recall rate of 60.4%, versus 61.5% for the quantized GLM-5.3. Over three consecutive test rounds, both models detected at least 23 vulnerabilities once, while the full-precision GLM-5.3 found 25. Strictly speaking, Altar-1 is currently a GLM-5.3 pruned specifically for security scenarios. Aikido clarified that no new security-specific capabilities were trained in this iteration; work on security workflow fine-tuning, tool integration, and long-range inference training is still ongoing. NIST’s CAISI previously rated the base GLM-5.3 as the open-source model with the strongest cybersecurity capabilities to date, though its overall performance lags behind current leading U.S. models by approximately four months.