Eight New AI Models, One Sovereign Machine

We just added eight new open models to Xinity. They run on-premise, on a single AI machine, with no data leaving your building. Here's what each one is good for, in plain words.

If you run AI in a regulated industry like finance, healthcare, legal, public sector or media, you already know the hard part isn't finding a capable model. It's finding a capable model you're allowed to use. One that runs on your own hardware, answers without an internet connection, and never sends a single token to someone else's cloud.

So we integrated eight of them at once.

They cover the full range, from a 300-million-parameter model small enough for almost any hardware up to a 119-billion-parameter heavyweight. Together they handle writing and fixing code, understanding images, powering search, translating between languages, and careful step-by-step reasoning. All deployed on a single desk-side AI box.

Let's walk through them.

Two things worth clearing up first

A quick note on the two questions people always ask.

"One machine" means deployed, not all at once. All eight models live on a single AI box. You run the one you need for the job in front of you. You're not firing up all eight at the same time. Think of it as a toolbox that fits on one desk, not eight engines running in parallel.

"Private" means the models run offline. Once a model is deployed, it needs no internet connection to answer. Your prompts, your documents, your data all stay on the machine. Downloading the models in the first place needs a connection, like installing any software, but after that you can pull the plug and it still works.

That's the whole point of sovereign AI: powerful models, on infrastructure you control, with nothing leaving your four walls.

The eight models, in plain words

Qwen3.6 (27B): coding and tasks

A strong coding assistant that can also read images and work through multi-step jobs on its own. Comes in two efficiency levels, so you can trade a little capability for a lot of speed when you need to.

Good for: writing and fixing code, automating tasks, reading screenshots.

Qwen-AgentWorld (35B): AI agents

An unusual one. It imitates digital environments like a terminal, a web browser or an Android phone, so AI agents can plan and rehearse their actions safely before doing anything for real. If you're building agents, this is the practice ground.

Good for: building and testing AI agents.

EmbeddingGemma (300M): search

The little one. It turns text into a kind of fingerprint so software can tell what relates to what. Tiny, fast, and fluent in over 100 languages, which makes it the quiet workhorse behind search and document organisation.

Good for: search, recommendations, organising documents.

Gemma 4 (compact): everyday assistant

A small, fast assistant that also understands images and uses tools. Nothing flashy, just a reliable generalist that runs comfortably on modest hardware.

Good for: quick everyday tasks on modest hardware.

VibeThinker (3B): reasoning

A small model that is very good at careful, show-your-work reasoning. When the steps matter as much as the answer, like in maths, logic or tricky code, this is the one that thinks it through rather than guessing.

Good for: math, logic, and coding problems where the steps matter.

Hunyuan MT2: translation

A dedicated translator covering 33 languages. Very small, very efficient, and focused on doing one thing well.

Good for: fast, high-quality translation.

LFM2.5: efficient assistant

An efficient assistant that switches on only a small part of itself for each word it generates, so it stays quick and light without giving up much capability.

Good for: an efficient general assistant that stays responsive.

Mistral Small 4 (119B): heavy lifting

The big one. A very large, very capable assistant that also understands images, compressed cleverly so it still fits on a single machine. When you need the most capability the box can hold, this is it.

Good for: heavier assistant and image tasks.

What stands out across the lineup

Big range, one box. Sizes run from small enough for almost any hardware up to a 119-billion-parameter giant, all deployed on a single AI machine.

Runs offline. Once deployed, the models need no internet connection to run. Your data stays on the machine.

Long memory. Several can take in the equivalent of hundreds of pages of text at once.

Multilingual. Several handle dozens, or even more than 100, languages out of the box.

Broad compatibility. They work on common, widely used setups, so you're not buying exotic hardware to get started.

Who this is for

If you're in a sector where the rules already decided that your data can't leave the building, this lineup is built for you. One machine, deployed on-premise, OpenAI-compatible, fully under your control, spanning code, images, search, reasoning and translation.

You don't have to choose one model and hope it's good at everything. You pick the right model for each job, and they all live on the same box.

That's sovereign AI the way it should work. Not a single model you compromise around, but the right tool for each task, on hardware you own.

Want to see the full lineup running on a single machine? Get in touch and we'll show you what sovereign AI looks like in practice.