Aikido unveils Altar-1: a pruned security model for enhanced deployment

Aikido Security has launched Altar-1, an open-weight security model derived from Z.AI’s GLM-5.3, reducing its size to 328 GB. This model is designed for deployment on customer-controlled infrastructure, enabling autonomous pentesting through Aikido Machine. Altar-1 is available on Hugging Face and runs on vLLM with a single node of 4x NVIDIA H200 GPUs. The model retains 92% of the original GLM-5.3’s covered vulnerabilities, with a KL divergence of 0.506 nats against the full BF16 version. Aikido tested Altar-1 on its internal CVE benchmark, identifying a critical-severity vulnerability during a client’s production pentest. The model inherits the GLM-5.3 license, allowing commercial use but requiring a Z.AI security review for operators with over $10B in revenue. Aikido plans to explore lower-bit formats and fine-tune models for security workflows. Altar-1 addresses deployment challenges by reducing memory usage, enabling a 128k-context KV cache on 4x H200 GPUs. It maintains routing efficiency by selecting 8 experts per token from 168, with about 40B active parameters. The model card specifies Hopper GPUs (H100 or H200) for optimal performance. Aikido also powers Aikido Attack, AI Code Analysis, and Deep Review, expanding its security toolset for enterprises and OT operators under data-residency mandates.