Skip to main content

A New ADITO-LLM, and a New Way to Version It

· 5 min read
Florian Brunner
Florian Brunner
Product Owner @ ADITO Platform
Tags:

ADITO-LLM has a new model: adito-llm-brizo. It is available today.

It is not the default yet. The bare adito-llm alias keeps pointing at the previous model until 31 August 2026. This post explains what changed, and why model names now work the way they do.

What brizo brings​

adito-llm-brizo runs on a newer base model: Qwen3.6, where adito-llm-athene runs Qwen3.5.

What that means in practice:

  • It follows instructions more closely, and structured output comes back in the shape you asked for.
  • Room for about 262,000 tokens in a single request, so long documents and long conversations fit without being split up first.
  • Faster responses, because the model is served in a more compact format (NVFP4 instead of FP8).

Thinking mode stays off by default, exactly as before. If you switch it on today, nothing about that changes.

Why you should move now​

Athene is deprecated. On 31 August 2026 it is retired, and the bare adito-llm alias flips to brizo on the same day.

First the reassuring part: a standard ADITO system does not break on that date. It ships configured against the model ID adito-llm, and that alias keeps working. On 31 August it resolves to brizo instead of athene, and your system carries on without a configuration change, an update, or anything to deploy. The same is true for the AI features in the product, the CKEditor AI Assistant included.

So if you do nothing, you get migrated automatically and everything keeps running. The only difference is that you meet the new model in production rather than in a test. Moving now gives you the rest of August to look at the thing that is going to happen anyway, on your own prompts and with time to adjust them.

Short notice this once​

The gap between brizo going GA and athene retiring is shorter than I would like it to be. That is a one-off. Athene was already running when we introduced the naming scheme described below, so this first transition had to fit around a model that existed before the rules did. From the next codename onwards, the transition window is published together with the model, and the sunset date is in the documentation long before it matters to you.

What backs that up is not goodwill. A model only reaches GA after it has passed our internal quality gates, and the alias only moves once the main stakeholders inside ADITO have signed off on the new model. Nothing switches underneath you because a new upstream release looked interesting on a Friday afternoon.

Two names, two promises​

Every model now has two kinds of name, and they make opposite promises.

  • adito-llm is a floating alias. It always points at the current model and upgrades on its own. You never have to do anything, and the model behind it will change.
  • adito-llm-brizo is a pinned name. It never changes. Everything that affects the output is frozen from its very first request: the base model, the settings we start it with, the context window, the built-in system prompt. If a model needs a fix, it gets the next name in the series instead of a quiet patch.

Names run alphabetically, so athene came before brizo and the next one starts with C. Each one moves through three stages: available, then deprecated, then retired.

None of this is specific to ADITO-LLM. Embeddings, speech-to-text and the code model use the same two kinds of name, with the same promise behind each. The code model is the one gap: it has an alias, but no pinned names yet.

One caveat, and it is a deliberate one: at retirement, everyone on the alias is migrated quietly, while anyone pinned to the retired name gets an error. That asymmetry is the point. Swapping a pinned model out from under you would defeat the reason to pin it in the first place.

While we are here: it reads images​

Almost nobody uses this, so it is worth saying out loud. ADITO-LLM does OCR. Put a scanned invoice or a photo of a delivery note into your message and the text comes back, on the same chat endpoint you already call. No second service, no second key. It is written up now under Image input and OCR.

Bigger models are coming​

Brizo is not the end of the size question. We are working on serving models considerably larger than what runs today, on the same API. More once it is ready.

Where to go next​

The current codenames, their status, and their sunset dates are in AI Models. Default request parameters and the full API reference are in Text Generation.

./Flo