Aikido Security has launched Altar-1, an open-weight security model pruned from Z.AI’s 753-parameter GLM-5.3 to 328 GB. Built to operate locally on customer infrastructure, Altar-1 powers Aikido Machine, an autonomous pentesting appliance designed for on-premise and air-gapped networks. The model can run on a single node equipped with four NVIDIA H200 GPUs using vLLM.
To solve data-residency and memory limitations without compromising specialist tasks, Aikido applied a two-step calibration and expert scoring process. The pruning reduced the model size by 78.2% compared to BF16 while preserving routing across 168 experts. On internal benchmarks covering 32 known vulnerabilities across 30 repositories, Altar-1 retained 92% of covered vulnerabilities compared to its parent model.
Altar-1 operates under the GLM-5.3 License, allowing commercial use and modification, though large Model-as-a-Service providers with over $10 billion in revenue require a security review. Aikido plans to explore lower-bit formats like EXL3 and fine-tune future models specifically for security workflows.
Why it matters
Enables enterprise security teams to run advanced autonomous pentesting models entirely on-premise, resolving strict data-residency constraints.
Demonstrates effective MoE pruning techniques, reducing model footprint to fit on a single 4x H200 GPU node while maintaining security domain performance.
Source: marktechpost.com



