Aikido Altar vs Massive AI Models: How 78.2% Compression Is Reshaping Cybersecurity Intelligence
Artificial intelligence is rapidly becoming part of the cybersecurity stack, helping security teams analyze code, identify vulnerabilities, investigate threats, generate remediation suggestions, and automate penetration testing. Yet the adoption of AI in security introduces a fundamental contradiction: the systems that need the most sensitive information are often the same systems that organizations are least willing to expose to external AI providers.
Belgian cybersecurity company Aikido is addressing that problem with Altar, an open-weight AI model designed specifically for defensive cybersecurity workloads. Introduced in September 2026, Altar is built around a straightforward proposition, security intelligence should be capable of running inside the infrastructure it is protecting.
The model represents a broader shift in cybersecurity AI, from relying primarily on remotely hosted general-purpose models toward specialized systems that can operate locally, including in environments with strict data-residency requirements or no internet connectivity.
Why Sovereign AI Matters in Cybersecurity
Security operations routinely involve information that organizations consider exceptionally sensitive. Source code, internal architecture, vulnerability reports, configuration files, business logic, credentials, application behavior, and unpatched security findings can all reveal how an organization works and where it is exposed.
Sending that information to an external inference service can therefore create governance and compliance challenges even when the underlying AI service has strong security controls.
The problem is particularly significant in sectors such as banking, healthcare, government, industrial infrastructure, and other environments subject to strict regulatory or operational requirements. Some organizations also operate isolated networks where connecting security tooling to an external AI service is simply not an option.
This has created growing interest in sovereign AI, meaning AI capabilities that can be deployed and controlled within an organization's own infrastructure.
Altar extends this concept into cybersecurity. Rather than requiring sensitive security context to leave an organization's environment, the model can be deployed locally and integrated with security workflows operating on-premises.
Aikido had already pursued this architecture through its autonomous penetration-testing appliance, Aikido Machine, which can operate entirely within customer infrastructure, including air-gapped environments. Altar adds a specialized AI reasoning layer to that model of deployment.
The significance extends beyond one product. It highlights a central question for enterprise AI, whether maximum model capability is always more valuable than operational control.
The Model Deployment Problem
The challenge with deploying advanced AI locally is not simply obtaining the model weights. Modern large language models can require enormous amounts of memory and computational infrastructure, particularly when they use mixture-of-experts architectures.
In a mixture-of-experts model, different neural-network components specialize in different patterns of computation. Only a subset of experts may be activated for a particular token, but the broader expert collection still has to be available to the serving infrastructure.
This creates an unusual efficiency problem.
A cybersecurity agent investigating a vulnerability might need strong programming knowledge, security reasoning, tool-use capabilities, and multilingual comprehension. It may not require every capability encoded across a huge general-purpose model. Nevertheless, a conventional deployment can still carry the memory burden of the complete model.
Agentic workloads make the issue more difficult.
An autonomous security investigation can maintain a large context containing source-code observations, previous tool calls, command results, hypotheses, vulnerability evidence, and intermediate reasoning. Multiple investigations running simultaneously compete for the same memory resources.
Consequently, reducing model size is not merely about lowering storage requirements. It can also create additional capacity for the long-running contexts that security agents need.
From 1.51 TB to 328 GB
Aikido built Altar from GLM-5.3, a large open model that the company identified as particularly strong in its internal security evaluations.
The full-precision model occupies approximately 1,506.7 GB. Quantization reduced that footprint to 488.2 GB, while subsequent expert pruning brought Altar down to 328.0 GB.
Model configuration | Stored model size |
GLM-5.3, BF16 | 1,506.7 GB |
GLM-5.3, AWQ INT4 | 488.2 GB |
Altar, pruned W4A16 | 328.0 GB |
The resulting Altar model represents a 78.2% reduction compared with the original full-precision model and a further 32.8% reduction compared with the already-quantized AWQ version.
Two different compression techniques are responsible for these reductions.
Quantization stores model parameters using fewer bits. In Altar's case, most expert weights are represented using four-bit values, while inference activations remain at 16-bit precision, producing a W4A16 configuration.
Expert pruning, by contrast, removes portions of the model itself. Altar retains 168 of the original 256 routed experts in each backbone expert layer, removing 88 experts, or 34.4%.
The distinction matters because compression is not automatically beneficial. Removing model capacity can also remove capabilities.
The technical challenge is therefore determining what can be discarded without disproportionately damaging the abilities required for cybersecurity.
Why Expert Selection Matters
Aikido used Cerebras REAP, or Router-weighted Expert Activation Pruning, to guide the selection of experts.
A simple strategy would be to remove experts based only on how frequently they are activated. That approach can overlook less common but important capabilities. An expert may contribute significantly to a specialized task even if that task represents a relatively small proportion of overall model activity.
Aikido instead examined contribution using information from the model's routing behavior and the magnitude of expert outputs. It also evaluated different groups of examples to reduce the possibility that specialized capabilities would disappear behind broad averages.
This is especially relevant to cybersecurity because security reasoning is not one isolated skill.
A capable security model may need to interpret programming languages, identify unusual application behavior, understand business requirements, reason about vulnerabilities, process technical documentation, and operate across multiple natural languages.
Aikido used traces from its internal penetration-testing harness as part of the selection process. These traces represented code, tool interactions, and model responses associated with security investigations. The company states that no customer data was used in this calibration process.
Multilingual material was also incorporated because security investigations may involve applications whose documentation, interfaces, workflows, or business rules are written in languages other than English.
The result is a key principle in specialized AI optimization: model size alone does not determine model usefulness. What survives compression can matter more than how much is removed.
Measuring the Security Trade-Off
Aikido evaluated Altar using an internal vulnerability benchmark containing 32 known vulnerabilities across 30 repositories, with three runs per case.
The results illustrate the trade-off created by compression.
Altar achieved an average recall of 60.4% per run and rediscovered 23 of the 32 vulnerabilities at least once across three runs.
The AWQ-quantized GLM-5.3 achieved 61.5% average recall and also covered 23 vulnerabilities. The original full-precision GLM-5.3 achieved 65.6% average recall and covered 25 vulnerabilities.
The comparison is important because it places the size reduction in context.
Against the AWQ model, Altar sacrificed roughly one percentage point of average recall while retaining the same vulnerability coverage in the benchmark. Compared with the original model, Altar retained 23 of the parent's 25 covered vulnerabilities while using substantially less storage.
However, these results should be interpreted within the boundaries of the benchmark. The evaluation focused on targeted vulnerability rediscovery within a larger pipeline. It was not designed to represent every stage of autonomous penetration testing, such as discovering previously unknown vulnerabilities across an entire codebase, executing exploits for validation, or proposing fixes.
That distinction illustrates why AI security benchmarks need to be understood in terms of their actual evaluation task rather than reduced to a single performance number.
From General-Purpose AI to Security-Native Models
Altar also illustrates an emerging architectural direction for enterprise AI.
General-purpose frontier models are designed to perform across an enormous range of tasks. That breadth is valuable, but it can impose substantial infrastructure costs when deployed for a narrowly defined workload.
Specialized models can take a different approach. Instead of maximizing universal capability, they can optimize the balance between capability, latency, memory consumption, privacy, and operational control for a particular domain.
Cybersecurity is particularly suitable for this approach because many workloads are highly technical and repetitive. Code analysis, vulnerability triage, security testing, remediation assistance, and threat investigation all involve recurring patterns that can be incorporated into specialized AI pipelines.
The future is therefore unlikely to be defined exclusively by the largest possible models. A heterogeneous AI environment could contain large general-purpose systems alongside smaller, domain-specific models selected according to the sensitivity and computational requirements of individual workloads.
Open-Weight AI and the Enterprise Security Market
The open-weight nature of Altar is another important component of its strategy.
Open-weight models provide organizations with greater control over deployment than conventional hosted AI services. They can potentially be integrated into private infrastructure, evaluated locally, adapted to specialized environments, and operated without transmitting sensitive workloads to an external inference provider.
That does not eliminate the operational burden. Running a 328 GB model still requires substantial hardware and expertise.
Aikido states that Altar can be deployed comfortably on a node equipped with four H200 GPUs using the latest version of vLLM. This makes the model substantially more practical for enterprise infrastructure than the original full-size checkpoint, while still placing it beyond the hardware requirements of ordinary business applications.
The economic equation is consequently broader than model licensing. Organizations must consider GPU infrastructure, power consumption, cooling, deployment engineering, model updates, monitoring, security controls, and integration with existing security systems.
The benefit is greater control over where security intelligence operates.
A Broader Shift Toward Local AI Security
The significance of Altar goes beyond compression.
Aikido describes a longer-term objective in which security models can improve through fine-tuning, better tool use, longer-horizon reasoning, and feedback from internal benchmarks. The company's roadmap also includes exploring lower-bit formats and additional serving optimizations.
This points toward a possible progression:
Compress general models to make local deployment practical.
Specialize them for cybersecurity through targeted training and evaluation.
Integrate them with security tools capable of taking actions.
Measure performance continuously using representative security benchmarks.
Improve future models using validated workflow data while maintaining privacy controls.
Such systems could increasingly operate as components of autonomous security platforms rather than simply acting as chat interfaces for security professionals.
The distinction is important. An AI that explains a vulnerability is useful. An AI that can inspect code, invoke security tools, reason over evidence, validate findings, and assist with remediation represents a much deeper transformation of the security workflow.
The Remaining Challenges
Local deployment does not automatically solve every cybersecurity AI problem.
Smaller or compressed models can lose capabilities. Security environments are also highly diverse, meaning a model optimized against one workload may perform differently against another. Language coverage, programming-language support, false positives, false negatives, tool reliability, and long-context behavior all remain important evaluation dimensions.
There is also a broader governance challenge. An autonomous security model needs appropriate permissions and safeguards because an AI system capable of security testing can potentially execute consequential actions if it is not properly constrained.
For these reasons, the most credible path is likely to involve AI operating within carefully designed security workflows, with explicit evaluation, access controls, auditability, and human oversight where appropriate.
Altar's deployment immediately into Aikido Machine also demonstrates why real-world validation matters. Aikido reported that shortly after deployment, the system identified a valid critical-severity vulnerability during a client production penetration test. That is a concrete operational result, but it should be viewed alongside controlled benchmark measurements rather than as a replacement for them.
The Future of Sovereign Security Intelligence
The deeper significance of Altar is the convergence of three technology trends: increasingly capable AI models, aggressive model optimization, and growing demand for data sovereignty.
For cybersecurity, these trends reinforce each other.
Organizations want AI that can reason over their most sensitive technical information. At the same time, they need systems that can operate within practical infrastructure constraints. Compression techniques such as quantization and expert pruning provide one path toward reconciling those requirements.
The result is a changing definition of AI capability. The most useful model for an organization may not be the largest model available. It may be the model that delivers sufficient reasoning quality while remaining affordable, deployable, controllable, auditable, and compatible with the environment it is intended to protect.
For technology analysts such as Dr. Shahid Masood and the expert team at 1950.ai, developments such as Altar illustrate a broader direction in AI and cybersecurity, where intelligence is increasingly moving from centralized services toward specialized systems embedded directly within the infrastructure of organizations.
The next stage of AI security will therefore be shaped not only by bigger models, but by better models, more efficient models, and models that can operate where sensitive data actually lives.
Key Takeaways
Aikido Altar is an open-weight AI model designed for defensive cybersecurity workloads and local deployment.
The model is derived from GLM-5.3 and combines quantization with expert pruning.
Its stored model size falls from approximately 1.51 TB to 328 GB, a 78.2% reduction from the full-precision model.
Altar retains 168 of 256 routed experts, removing 88 experts while preserving the model's core security capabilities.
On Aikido's internal CVE benchmark, Altar achieved 60.4% average recall and rediscovered 23 of 32 vulnerabilities at least once across three runs.
Local deployment can help organizations keep sensitive source code, architecture information, and security findings within controlled infrastructure.
The model demonstrates the growing importance of sovereign AI, particularly for regulated and air-gapped environments.
Future development is expected to move beyond compression toward security-specific training, stronger tool use, longer-horizon reasoning, and continuous evaluation.
The broader industry trend is toward specialized AI systems that balance intelligence with privacy, infrastructure efficiency, and operational control.
Comments