Aikido Releases 'Altar', a Lightweight Open-Weight Model for On-Premises Security

Aikido has released Altar, an open-weight security model that can run with full control (including air-gapped setups) within customer infrastructure. It overcomes the limitations of existing closed models, which require sending sensitive data such as source code or internal documents externally, and provides frontier-grade security defense capabilities even in banks, hospitals, and industrial OT environments with strict data residency regulations.

Model Lightweighting and Optimization

Altar applies Expert-pruning and Quantization to its base model, GLM-5.3, to lower deployment barriers. The size was reduced from the original 1.51 TB to 488 GB after quantization, and finally to 328 GB after pruning, achieving a total capacity reduction of 78.2%. This result was achieved by adopting the REAP method to retain only 168 of the 256 routed experts and applying the W4A16 quantization technique.

Performance and Deployment Specifications

In a benchmark targeting 32 vulnerabilities, Altar showed an average recall drop of 5.2 percentage points compared to the original model, but maintained vulnerability coverage at the 92% level. Notably, it further reduced storage capacity by 32.8% compared to already quantized models while minimizing performance degradation. Altar can be served smoothly on 4-H200 GPU nodes with the latest vLLM environment, and model weights are released via HuggingFace.