> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Aikido releases Altar-1, an open-weight security model built from GLM-5.3
- URL: https://globalfeed.ai/en/aikido-releases-altar-1-an-open-security-model-built-from-glm-5-3/
- Published: 2026-09-23T04:13:59.000Z
- Updated: 2026-09-23T14:29:37.000Z
- Description: Security company Aikido pruned and compressed Z.ai's open model GLM-5.3 to create Altar-1. It is designed to let organisations run security analysis on their own servers without sending source code outside.
- Author: GlobalFeed Editor
- Tags: Aikido, cybersecurity, open models, GLM, x-AikidoSecurity, dil-en, elle

Belgian application security company **Aikido** has released its first open-weight security model, **Altar-1**, on Hugging Face. The model is derived from Z.ai's open model GLM-5.3\. Aikido pruned and compressed GLM-5.3 for security work, bringing it down to a size organisations can run on their own servers.

## Why it was built

According to Aikido, banks, hospitals and industrial operators on isolated networks cannot send their source code and open findings to cloud AI services. The company says it built Altar so these organisations can run security analysis entirely inside their own infrastructure. The model also powers Aikido Machine, the company's autonomous pentesting appliance that runs inside the customer's network.

## How it was shrunk

GLM-5.3 is a very large mixture-of-experts (MoE) model. Each layer has many small expert networks; only a few run for each token, but all of them must be kept in memory. Aikido used two methods:

- **Quantisation:** Expert weights are stored in 4 bits instead of 16\. This step started from an existing community 4-bit build of GLM-5.3.
- **Expert pruning:** Using Cerebras's REAP method, the experts contributing least to security work were removed. Aikido decided which experts to keep using traces from its own pentesting harness, coding and tool-use examples, and multilingual text. Aikido says no customer data was used.

The result keeps 168 of GLM-5.3's 256 experts per layer and comes to 328 GB. According to the table in Aikido's blog, full-precision GLM-5.3 takes about 1.5 TB and the quantised-only version 488 GB. The pruning itself was carried out by community developer 0xSero. Aikido credits Z.ai, Cerebras and the contributing developers on the model card.

## How much is lost

Aikido measured the model on its own benchmark of 32 known vulnerabilities across 30 repositories. According to the company, Altar found 23 of them at least once. Full GLM-5.3 found 25 on the same test, and the quantised-only version the same 23 as Altar. Aikido notes that the test measures rediscovery of known vulnerabilities, and does not cover blind discovery across a whole codebase or validating findings by exploiting them.

## How to use it

The weights can be downloaded from the AikidoSec/altar-1 page on Hugging Face, and the model runs with vLLM. Aikido also offers help with on-premises deployment. The company's Aikido Attack, AI Code Analysis and Deep Review products already use Altar.

## Things to keep in mind

- "Local" here means the organisation's own data centre. The model is prepared for Hopper-generation data-centre GPUs, such as four NVIDIA H200s. It is sized for server hardware, beyond a personal computer.
- The model has not been retrained; it carries GLM-5.3's abilities. Aikido plans fine-tuning for security work as a next step.
- The license is inherited from GLM-5.3\. Read that license before commercial use.
- For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Aikido blog: Introducing Aikido Altar](https://www.aikido.dev/blog/aikido-altar-open-weight-ai-sovereign-security?ref=globalfeed.ai), [Hugging Face: AikidoSec/altar-1 model card](https://huggingface.co/AikidoSec/altar-1?ref=globalfeed.ai), [REAP paper (Cerebras)](https://arxiv.org/abs/2510.13999?ref=globalfeed.ai).