Aikido releases open-weight security LLM Altar-1, built on GLM-5.3

Belgian application-security company Aikido has released Altar-1, an open-weight large language model built for on-premises and air-gapped security work such as vulnerability triage and code review. The model is a compressed derivative of Z.AI's 753-billion-parameter GLM-5.3, and its weights are published on Hugging Face — but it ships under GLM-5.3's license, not a permissive open-source one, which matters for anyone planning to redistribute or resell it.

How Altar-1 was built

Aikido started from an INT4-quantized version of GLM-5.3 and applied expert pruning — removing 88 of the model's 256 routed experts per layer — using Cerebras Research's REAP method. The result is a 328GB model, down from roughly 1.5TB for the full-precision original, that Aikido says can be served on four Nvidia H200 GPUs with vLLM. On an internal benchmark of 32 known vulnerabilities across 30 repositories, Aikido reports Altar-1 identified 23 of them, versus 25 for the unpruned model — a claimed retention of about 92% of the larger model's detection coverage.

What it is for, and what it is not

Aikido positions Altar-1 as the engine behind its autonomous penetration-testing product, Aikido Machine, designed to run entirely inside a customer's own network, including fully air-gapped environments, so source code and unpatched findings never leave the customer's infrastructure. That framing is specific: it is a defensive triage and code-review tool, not a general-purpose exploit generator, and Aikido's benchmark is an internal test rather than an independently audited one.

What buyers should check before deploying

Two practical points apply to any MSP or enterprise evaluating this class of tool. First, Altar-1 inherits GLM-5.3's license terms rather than MIT or Apache — that needs a legal review before installing it for a customer, especially in air-gapped or resold deployments. Second, Aikido's coverage numbers come from its own test set; treat them as a vendor claim pending third-party verification, the way any benchmark from a model's own maker should be read.

For teams weighing self-hosted security assistants, Altar-1 is a concrete example of the current pattern — pruned, quantized versions of large open-weight models tuned for a narrow defensive task — rather than a signal that any such tool should run unsupervised against production systems.

Sources: Aikido — Introducing Altar-1, Hugging Face — AikidoSec/altar-1, Security Boulevard coverage.