Aikido introduced Altar on September 21, describing it as an open-weight model for defensive security work in infrastructure controlled by the customer.
The company’s technical announcement says Altar is based on GLM-5.3 and supports its Aikido Machine offering. The intended use includes organizations that cannot send sensitive code and security findings to an external inference service.
Deployment size remains substantial
Aikido reports reducing the model’s storage footprint from 1.51 TB at full size to 488 GB through quantization and then to 328 GB through expert pruning. Those figures and the associated quality claims are supplied by Aikido.
The smaller version is still an infrastructure-scale deployment. Model size also does not describe the full memory and compute requirement under a particular workload.
For a security team, the useful comparison includes throughput, supported context and the quality of findings on its own systems. An available model is only one component of an operational security workflow.
Open weights and local operation answer different questions
“Open-weight” describes availability of model parameters; it does not by itself establish unrestricted licensing or complete openness of the training process. The release terms need a separate review.
Local deployment addresses where inference happens. It does not prove that every finding is valid or that an automated test is safe to run against a production service.
Our open-weight and open-source explainer sets out the distinctions. A controlled evaluation should verify findings against independent evidence and keep the system’s permitted targets explicit.
