Altar-1: Aikido Security’s Open-Weight AI Model Revolutionizes On-Premise Security

Bridging the Gap: AI Security for Sensitive Environments

At a glance, In today’s complex cybersecurity landscape, organizations face a critical dilemma: leveraging the power of advanced AI for security while safeguarding highly sensitive data. Traditional frontier AI models often operate on third-party infrastructure, forcing critical security context—like source code, architectural documents, and unaddressed vulnerabilities—outside the controlled network. This poses significant challenges, particularly for industries under strict data-residency mandates, such as banking, or for operational technology (OT) environments with no internet connectivity.

Meanwhile, Aikido Security has stepped forward with an innovative solution: Altar-1. This is their inaugural open-weight security model, meticulously engineered to run directly within a customer’s own infrastructure. Altar-1 is a highly compressed version of Z.AI’s powerful GLM-5.3 model, designed to empower autonomous pentesting and advanced security analysis without compromising data sovereignty.

The Core Problem: Data Residency and Deployment Challenges

The need for AI-driven security is undeniable, but the methods for deploying these powerful tools have been fraught with issues:

• Closed Frontier Models: Using these means sending proprietary and sensitive data outside the network to be processed on external servers, violating data residency requirements and increasing risk.
• Open-Weight Models’ Deployment Gap: While open-weight models solve the data residency problem by allowing on-premise deployment, large Mixture-of-Experts (MoE) models can be prohibitively large. They require storing every expert, even if only a few are active for a given workload. Furthermore, the substantial KV (Key-Value) cache built by security agents during long-running contexts competes fiercely with model weights for precious GPU memory.

In practical terms, Altar-1 directly addresses these hurdles, offering a viable path for organizations to harness cutting-edge AI security within their secure perimeters.

Introducing Altar-1: A Lean, Powerful Security AI

Altar-1 is not just another AI model; it’s a strategically optimized tool built for practical, secure deployment. It currently powers Aikido Machine, the company’s autonomous pentesting appliance, making it ideal for both on-premise and air-gapped networks.

How Altar-1 Was Built: Smart Compression

For example, The original GLM-5.3 is a massive 753-billion parameter MoE model. To make it deployable in real-world, resource-constrained environments, Aikido Security applied two sophisticated compression steps:

1. Step 1: Quantization

   Altar-1 begins with the cyankiwi GLM-5.3-AWQ-INT4 checkpoint. This involves applying AWQ (Activation-aware Weight Quantization) to store routed expert weights in 4 bits (W4A16 format), while maintaining 16-bit activations. Crucially, critical components like attention mechanisms, shared experts, dense layers, and the model’s head remain in BF16 for optimal performance.

2. Step 2: Expert Pruning

   That said, Aikido utilized Cerebras REAP (Router-weighted Expert Activation Pruning). Unlike simpler pruning methods, REAP intelligently scores each expert based on its router weight and output magnitude, not just its selection frequency.

   This allowed Altar-1 to retain 168 out of 256 routed experts per layer, effectively removing 88 experts (a 34.4% reduction) without requiring any retraining. This process ensures that specialist experts vital for code analysis, rare languages, and structured output are preserved.

The model’s calibration process rigorously used traces from Aikido’s pentesting harness, alongside diverse data like coding, tool calling, reasoning, and multilingual Wikipedia text. Aikido emphasizes that no customer data was used in this calibration, further reinforcing its commitment to privacy and security.

Dramatic Size Reduction

The results of this meticulous compression are striking:

• GLM-5.3 (BF16): 1,506.7 GB
• GLM-5.3 (AWQ INT4): 488.2 GB
• Altar-1 (Pruned W4A16): 328.0 GB

This represents a staggering 78.2% size reduction compared to the full BF16 GLM-5.3 and a 32.8% reduction from its AWQ parent. Despite this massive compression, Altar-1 maintains impressive fidelity, showing a KL divergence of just 0.506 nats against the full BF16 model on a sealed 25-prompt panel.

Performance Benchmarks and Real-World Impact

However, Aikido Security rigorously tested Altar-1 on its internal CVE benchmark, which covers 32 known vulnerabilities across 30 repositories. While the benchmark’s scope is specific to targeted CVE rediscovery within a larger pipeline, the results are promising:

• GLM-5.3 (BF16): 65.6% average recall, found 25 of 32 vulnerabilities.
• GLM-5.3 (AWQ INT4): 61.5% average recall, found 23 of 32 vulnerabilities.
• Altar-1: 60.4% average recall, found 23 of 32 vulnerabilities.

Compared to the AWQ checkpoint, pruning resulted in only about a 1-point recall drop with no loss in coverage. Against the original parent model, Altar-1 retained 92% of covered vulnerabilities (23 of 25) with a 5.2-point lower recall. Beyond benchmarks, Aikido also reports a successful real-world finding: Altar-1 identified a critical-severity vulnerability during a client’s production pentest.

Deployment and Licensing

Meanwhile, Deploying Altar-1 requires robust hardware: a single node equipped with 4x NVIDIA H200 GPUs. This configuration provides the 328 GB needed for the model weights, leaving ample room for a 128k-context KV cache at production batch sizes. It runs efficiently with vLLM, which automatically selects the Marlin MoE kernel.

Altar-1 operates under the GLM-5.3 License, which permits commercial use, modification, and redistribution. However, it’s important to note a specific clause: Model-as-a-Service operators generating over $10 billion in revenue within a 12-month period must undergo a security review by Z.AI. It’s crucial to understand that Altar-1 is an open-weight model, not an OSI-approved open-source solution.

In practical terms, Beyond its role in Aikido Machine, Altar-1 also powers Aikido Attack, AI Code Analysis, and Deep Review, showcasing its versatility across different security workflows.

The Road Ahead

Aikido Security’s plans for Altar-1 include exploring even lower-bit formats, such as EXL3, to potentially retain more experts while further reducing size. They also intend to fine-tune these models specifically for various security workflows, promising even more specialized and effective AI-driven defense mechanisms in the future.

Key Takeaways for Security Professionals:

• Massive Compression: Altar-1 shrinks GLM-5.3 from 1,506.7 GB to a deployable 328 GB (a 78.2% reduction).
• Intelligent Pruning: Utilizes REAP to keep 168 of 256 experts per layer, ensuring critical functionality.
• Strong Performance: Maintains 92% of the parent model’s CVE coverage with a manageable recall drop.
• On-Premise Ready: Designed to run on 4x H200 GPUs with vLLM, ideal for air-gapped and data-sensitive environments.
• Commercial Use: The GLM-5.3 License allows commercial application, with specific terms for very large enterprises.

Expert Perspective

A practical read on Altar-1 security model starts with security. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make Altar-1 security model a meaningful reference point across altar.

For decision-makers, the useful lens is not the headline alone but how data changes priorities once organizations have to respond.

Frequently Asked Questions

Why is Altar-1 security model important?

Bridging the Gap: AI Security for Sensitive EnvironmentsAt a glance, In today’s complex cybersecurity landscape, organizations face a critical dilemma: leveraging the power of advanced AI for security while safeguarding highly sensitive data.

What impact could Altar-1 security model have?

Traditional frontier AI models often operate on third-party infrastructure, forcing critical security context—like source code, architectural documents, and unaddressed vulnerabilities—outside the controlled network.

What should readers watch next with Altar-1 security model?

This poses significant challenges, particularly for industries under strict data-residency mandates, such as banking, or for operational technology (OT) environments with no internet connectivity.Meanwhile, Aikido Security has stepped forward with an innovative solution: Altar-1.

How does this relate to security?

It connects because the article frames security as one of the clearest areas where the topic may be felt in practice.

Source: https://www.marktechpost.com/2026/09/25/aikido-security-releases-altar-1-an-open-weight-security-model-pruned-from-glm-5-3-to-328-gb/