Aikido shrinks a security model to run on your own servers

Aikido has described Altar, its first open-weight security model, which the AI Search newsletter reported on 2026-09-27. According to aikido.dev, Altar is built to bring defensive security capability into infrastructure the customer controls, so an organisation's most sensitive context does not go to a third-party inference service.

What Aikido says Altar is

Aikido states that it started from GLM-5.3, which it calls one of the strongest models in its security evaluations. It reports a total 78.2% size reduction against that model at full precision, and 32.8% against the same model with AWQ INT4 quantization. The weights are published under Aikido's organisation, and Aikido says Altar can be deployed and served using a 4-H200s node with the latest version of vLLM, including in fully air-gapped environments.

Why the size number matters more than the model

The interesting part for anyone running AI on company work is not the security use case itself but the pattern. Aikido took a model it judged strong, removed part of it, and kept the result inside the customer's own infrastructure. That is the same constraint operations, finance and procurement teams meet when they want a model to read their invoices, contracts or internal documents without shipping them to an outside service.

The limitation is that the reported figures are Aikido's own, covering its own compression work and its own evaluation setup. The reduction is real in the sense that the storage numbers are stated; whether the smaller model behaves well on your documents is a separate question that only your own material can answer.

For the work BARGO automates, the closest fit is Document Desk and Company Brain: reading long, messy documents and answering questions from a company's own files with the source passage. A team that wants to test this direction would start with a small, representative set of its own documents, run the same extraction or question-answering task through a compressed model and a larger one, and compare the answers side by side. Approval Runner raises a related question, since routing decisions need consistent structured output rather than fluent prose.

Aikido's own framing is that defenders should be able to run these models inside the environments they protect. The same argument applies to the paperwork a company would rather not send anywhere.