Aikido Launches Altar-1, a Compressed GLM-5.3 Model for On-Premise Security Use

ME News report, September 22 (UTC+8), Beating AI Bulletin: Belgian cybersecurity company Aikido has released its first open-source security model, Altar-1. It is a compressed version of Z.ai’s GLM-5.3, optimized for security use cases. Enterprises can deploy it on their own servers—even in air-gapped environments—for code auditing and penetration testing, without exposing source code or unpatched vulnerabilities to external models. Aikido first quantized GLM-5.3 to INT4, then pruned 88 of the 256 routing experts using Cerebras’ REAP method, reducing capacity by 34.4%. The model size shrank from 1.51 TB in full precision to 328 GB—32.8% smaller than the quantized GLM-5.3’s 488 GB. The official recommendation is deployment using four H200 GPUs. Performance remains largely intact after compression. In Aikido’s proprietary benchmark of 32 CVEs, Altar-1 achieved an average single-round recall rate of 60.4%, compared to 61.5% for the quantized GLM-5.3. Over three consecutive rounds, both models detected at least 23 vulnerabilities each, while the full-precision GLM-5.3 detected 25. Strictly speaking, Altar-1 is currently more accurately described as “GLM-5.3 with security-focused pruning.” Aikido explicitly states that no new capabilities were trained in this phase; true fine-tuning for security workflows, tool use, and long-horizon reasoning remains forthcoming. NIST’s CAISI previously rated the base GLM-5.3 as the most capable open-source model for cybersecurity today, though its overall performance still lags behind leading U.S. models by approximately four months. (Source: BlockBeats)