Altar-1 compresses GLM-5.3 to run security AI locally

With Altar-1, Aikido adapts GLM-5.3 for cybersecurity environments that cannot send their code or sensitive data to an external service. The open-weight model shrinks from 1.51 TB to 328 GB while retaining, according to Aikido's internal testing, most of the measured vulnerability detection capabilities.

Keeping the model with the data

For banks, hospitals, or industrial environments isolated from the internet, sending source code, internal architecture documents, or unpatched vulnerabilities to an external API may simply not be an option. That constraint sits at the center of Altar-1, Aikido's first open-weight security model, designed to run inside an organization's own infrastructure, including fully air-gapped environments.

The model also powers Aikido Machine, the company's autonomous pentesting system. Instead of moving sensitive information to a remotely hosted model, the approach brings the analysis capabilities into the environment where the code and data already reside. From 1.51 TB to 328 GB

Altar-1 starts with GLM-5.3. The full BF16 model takes up roughly 1,506.7 GB. AWQ INT4 quantization first brings that down to 488.2 GB, before expert pruning reduces the stored weights to 328 GB. That represents a 78.2% reduction compared with the full-precision model and 32.8% compared with its already quantized version.

GLM-5.3 uses a Mixture of Experts architecture. The complete expert pool still needs to be stored even though only a subset is activated for each token. Altar keeps 168 of the original 256 routed experts and removes 88. Eight experts are still selected per token, but from a smaller pool.

Aikido used Cerebras' REAP pruning method to decide what remained. Selection was calibrated using traces from the company's pentesting harness, including code and tool calls, alongside multilingual text intended to preserve the model's ability to understand documentation and interfaces in different languages. Aikido says no customer data was used during this process.

This does not yet amount to teaching the model new cybersecurity skills through fine-tuning. Aikido lists that as a future step, alongside work on tool use and long-horizon reasoning. Testing against 32 known vulnerabilities

Aikido evaluated the compressed model on an internal benchmark containing 32 known CVEs across 30 repositories, with three runs for each case. Altar-1 averaged 60.4% recall per run and rediscovered 23 of the 32 vulnerabilities at least once.

The AWQ INT4 version of GLM-5.3 reached 61.5% and covered the same 23 vulnerabilities. Full-precision GLM-5.3 reached 65.6% and covered 25. Altar therefore retained 23 of the 25 vulnerabilities found by its full parent at least once, with a 5.2 percentage-point decrease in average recall.

Those numbers have a deliberately narrow scope. Aikido's benchmark measures targeted rediscovery of known vulnerabilities inside its own pipeline. It does not test blind discovery across an entire repository, exploit execution to validate findings, or the quality of proposed fixes. No independent replication of these results has been published yet. Four H200s to run it

The Altar-1 weights on Hugging Face occupy 328 GB and are distributed under the license inherited from GLM-5.3. Aikido's deployment configuration serves the model through vLLM on a system equipped with four H200 GPUs.

Aikido says Altar has already been deployed across its Aikido Machine fleet and subsequently identified a valid critical-severity vulnerability during a client production pentest. For now, that result remains a company-reported case rather than an independently documented evaluation.