Belgium-based cybersecurity company Aikido has released Altar-1, its first open-source security model. The model is a compressed version of Z.ai’s GLM-5.3 tailored for security use cases.
Aikido said enterprises can deploy it on their own servers, including fully air-gapped environments, for code auditing and penetration testing. That setup allows source code and unpatched vulnerabilities to stay off external models.
How Aikido compressed the model
Aikido first converted GLM-5.3 into an INT4-quantized version. It then applied Cerebras’ REAP method and pruned 88 of the model’s 256 routing experts, equal to 34.4%.
The model size fell from 1.51 TB in full precision to 328 GB. That is also 32.8% smaller than the 488 GB quantized version. Aikido’s official recommendation is to deploy Altar-1 on four H200 GPUs.
Benchmark results
According to Aikido’s self-built benchmark covering 32 CVE tests, Altar-1 recorded an average single-run recall rate of 60.4%. The quantized GLM-5.3 posted 61.5%.
Across three consecutive test runs, both models found 23 vulnerabilities at least once, while the full-precision GLM-5.3 found 25. Aikido said performance was largely preserved after compression.
The company also said that, strictly speaking, Altar-1 is currently closer to a security-oriented pruned version of GLM-5.3.
What comes next
Aikido said this step did not train any new skills into the model. Fine-tuning for real security workflows, tool calling and long-horizon reasoning training are still ahead.
NIST-affiliated CAISI had previously rated the base GLM-5.3 as the strongest open-source model for cybersecurity capability at present, though its overall capability still trails current leading U.S. frontier models by about four months.