Aikido Security has released Altar-1, an open-weight security model built by compressing Z.AI's GLM-5.3 down to 328 gigabytes. The Belgian cybersecurity firm says the model keeps 168 of the original 256 routed experts per layer and uses 4-bit weight, 16-bit activation quantisation to cut the footprint from roughly 1.5 terabytes.
The pruning process relied on traces from Aikido's own penetration-testing harness and multilingual text, not customer data, according to the company. Altar-1 powers Aikido Machine, an on-premises pentesting appliance aimed at organisations that cannot send code or infrastructure details to an external API. It runs on a single node with four Nvidia H200 GPUs via vLLM or SGLang, including in air-gapped environments.
Reported benchmark figures vary between sources: one puts retained vulnerability coverage at around 92% of the full model, while another cites 60.4% recall on a 32-vulnerability internal CVE benchmark versus 65.6% for the unpruned model. Aikido has not published a single definitive figure, so the numbers should be treated as a range rather than a fixed claim.
Belgian bank Belfius is among the customers expected to use Altar-1 through Aikido's existing products. Shortly after release, the company said the model flagged a critical-severity vulnerability during a live penetration test. Aikido Labs is now looking at further compression, including lower-bit formats such as EXL3, and fine-tuning the model specifically for security workflows.
Source: MarkTechPost
