AI Models

The ADITO AI Runtime serves several kinds of models behind one self-hosted platform: text generation, code generation, embeddings, and speech-to-text. Text generation, embeddings, and speech-to-text use an OpenAI-compatible API. Code generation supports whichever protocol the coding agent expects — OpenAI-compatible, Anthropic Messages, or OpenAI Responses; see Code Generation. All model kinds share the same infrastructure, authentication, and data-sovereignty guarantees described in AI Compliance.

Model kinds

Kind	Purpose	Model	Page
Text Generation	Summarize, translate, rewrite, and classify text; read text from images (OCR)	adito-llm	Text Generation
Code Generation	Coding agents (OpenCode, Claude Code, Codex)	adito-code	Code Generation
Embeddings	Semantic and similarity search	adito-embed	Embeddings
Speech-to-Text	Transcribe audio to text	adito-stt	Speech-to-Text

"Model" is the floating alias; configure this unless you need a fixed, unchanging version. See Model naming and versioning below for how pinning works.

Each model page follows the same structure: a model card with the technical details, example use cases, and a first how-to to get started. Model kinds with more than one pinned codename also list a "Pinned versions" table with each codename's status and sunset/retired date.

Request limits

Two limits apply to every request, regardless of the model kind:

Limit	Current value	What it covers
Request timeout	120 seconds	Wall-clock time for a single request. A request that has not finished by then is aborted.
Maximum request size	20 MB	The complete request body, including base64-encoded images and uploaded audio files.

Both are current platform settings rather than properties of a model. Unlike sampling defaults and the context window, they are not frozen by a pinned codename and can change without a new codename being issued.

Practical consequences:

• Split long work into several requests instead of pushing a single one to the limit. Thinking mode, long documents, and a high max_tokens value together are the usual cause of a timeout. See Thinking mode.
• Base64 encoding inflates a payload by about a third, so an image or audio file should stay well below 20 MB before encoding. Send pages or recordings one at a time rather than bundling them.

Model naming and versioning

Model identifiers such as adito-llm are ADITO's own naming — the underlying upstream model is documented on each model's page but never appears in the API identifier.

One versioning scheme covers the whole runtime. Text generation, code generation, embeddings, and speech-to-text all follow the rules below; only the codenames and dates differ.

Each model kind exposes two identifier types:

• A floating alias (e.g. adito-llm) always points to the current GA model and upgrades automatically. This is the default used throughout this documentation.
• A pinned codename (e.g. adito-llm-athene) is an immutable version — base model, quantization, sampling defaults, context window, and system prompt are frozen. Codenames are assigned alphabetically as new versions ship, and each one moves through Active (GA) → Deprecated → Retired.

Because sampling defaults and the context window are part of what a codename freezes, they are documented per model kind rather than here: see the "Default request parameters" section on each model page.

The same rules for every model

Identifier naming, the immutability contract and the Active (GA) → Deprecated → Retired lifecycle are the same for all four model kinds. No model kind has its own versioning rules. The only difference today is whether pinning exists yet: adito-code is served through its floating alias only, while text generation, embeddings and speech-to-text also expose pinned codenames.

How a model change rolls out

A new codename ships as Active (GA) right away, so you can pin it and start tuning your integration early. The floating alias keeps pointing at the previous, now-deprecated codename during a transition window. Alias and old codename flip together at retirement — pin early to get ahead of the switch, or do nothing and get migrated automatically.

Current codenames by model kind

Model Task	Floating Alias	ModelID	Base Model	Status	Sunset / Retired Date
Text Generation	adito-llm	adito-llm-athene	Qwen3.5-35B-A3B-FP8	Deprecated	2026-08-31
Text Generation	— (pin only, not yet default)	adito-llm-brizo	Qwen3.6-35B-A3B-NVFP4-Fast	Active (GA)	—
Code Generation	adito-code	— (not pinnable)	Qwen3.6-35B-A3B	Active (GA)	—
Embeddings	adito-embed	adito-embed-andromeda	Qwen/Qwen3-Embedding-0.6B	Active (GA)	—
Speech-to-Text	adito-stt	adito-stt-accordion	Whisper large-v3	Active (GA)	—

Pinning isn't available yet for Code Generation (adito-code) — floating alias only. Use the floating alias unless your integration needs guaranteed, unchanging output over time.

---

See also: Text Generation | AI Compliance | AI Overview