Aikido Security Releases Altar-1 Security Model
Aikido Security launches Altar-1, a 328 GB open-weight security model pruned from GLM-5.3, deployable on a node of 4x NVIDIA H200 GPUs.
Stock photo for illustration only, not from the actual event
- Aikido Security launches Altar-1, a 328 GB open-weight security model.
- Pruned from GLM-5.3 to solve GPU memory bottlenecks and data residency issues.
- Runs on a node of 4x NVIDIA H200 GPUs with support for a 128k context KV cache.
- Model weights are now publicly available on Hugging Face for vLLM deployment.
Aikido Security has officially released Altar-1, a 328 GB open-weight security model pruned down from the foundational GLM-5.3 model. The new release aims to solve deployment bottlenecks for organizations bound by strict data-residency mandates, such as financial institutions and operational technology (OT) operators that operate without internet routes and cannot send source code or architecture documents to external cloud infrastructures.
Traditional Mixture-of-Experts (MoE) models like GLM-5.3 feature 753B parameters and require storing every expert in memory, even when workloads utilize only a fraction of them. Furthermore, security agents generate long-running context, causing the KV cache to compete directly with model weights for the exact same GPU memory resources, complicating production deployments.
Stock photo for illustration only, not from the actual event
The Aikido team applied two compression steps during development. Calibration utilized traces from their internal pentesting harness, alongside coding, tool calling, reasoning, and multilingual Wikipedia text, while confirming no customer data was utilized. Each expert was scored by its largest share of routed work across domains, protecting specialists dedicated to code, rare languages, and structured outputs.
Pruning a massive MoE architecture like GLM-5.3 down to 328 GB represents a crucial milestone in bringing advanced AI capabilities directly to on-premises cybersecurity environments. Preserving domain-specific expert routing ensures that the model maintains high vulnerability detection accuracy without forcing enterprises to compromise data privacy by relying on external APIs.
Benchmark results show that Altar-1 is 78.2% smaller than BF16 and 32.8% smaller than the AWQ parent. In terms of fidelity, Altar-1 achieves a KL divergence of 0.506 nats against full BF16 across a sealed 25-prompt panel. When evaluated on Aikido's internal CVE benchmark covering 32 known vulnerabilities across 30 repositories, the pruning incurred minimal performance loss.
Hardware requirements specify Hopper architecture GPUs (H100 or H200). Utilizing 328 GB across four H200 GPUs leaves sufficient overhead for a 128k-context KV cache at standard production batch sizes. Altar-1 inherits the GLM-5.3 license, permitting commercial use, modification, and redistribution, though Model-as-a-Service operators generating over $10 billion in revenue over 12 months must pass a Z.AI security review.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment