The AION catalog.Running in one click.
Every model is the version quantized and optimized by AION. Download, start, stop; the full stack starts in one go.
AION stack
5 connected models, one click to start them all.
- LLMQwen 3.8 Flash Next140BDownloaded
- EmbeddingsQwen3 EmbeddingDownloaded
- RerankerQwen3 RerankerDownloaded
- OCRGLM-OCRDownloaded
- STTWhisper Large v3Downloaded
| Role | AION model | Version | PGX | State | Action |
|---|---|---|---|---|---|
| Qwen 3.8 Flash Next140BLLM · NVFP4 · AION · 1 PGXDownloaded | Downloaded | ||||
| Qwen3 EmbeddingEmbeddings · AION · 1 PGXDownloaded | Downloaded | ||||
| Qwen3 RerankerReranker · AION · 1 PGXDownloaded | Downloaded | ||||
| GLM-OCROCR · AION · 1 PGXDownloaded | Downloaded | ||||
| Whisper Large v3STT · AION · 1 PGXDownloaded | Downloaded | ||||
| GLM 5.3753BLLM · AION · 2 PGX58%Downloading | 58%Downloading | ||||
| Qwen4LLM · AION · 1 PGXDownloaded | Downloaded | ||||
| Qwen3.635B A3BLLM · PQv2 · AION · 1 PGXIn the catalog | In the catalog |
One PGX runs models up to 140 billion parameters; with two linked PGX you reach GLM 5.3, at 753 billion. Try the board: download a model, start it, stop it. The catalog holds no original weights: every entry is the version AION has quantized and optimized for the machine.
The whole stack.Or just one model.
Hub starts models the way today's work needs them: the whole stack that Agents and document search rely on, or just the model you picked.
The full stack
One click starts all five roles together. They come up already connected, in the right order, with nothing to configure by hand.
- LLM
- Reasons, writes and answers.
- Embeddings
- Searches by meaning.
- Reranker
- Puts the results in order.
- OCR
- Reads scans and images.
- STT
- Turns speech into text.
They all answer from a single OpenAI-compatible gateway: Agents and the tools you already use find them at the same address.
A single model
Need one more model, or a different one? Pick it from the catalog and manage it on its own.
- 01
Pick
Browse the whole catalog and filter by role.
- 02
Download
The model reaches the machine once.
- 03
Start
One click and it answers from the gateway.
- 04
Stop
When you don't need it, stop it and free memory.
Not the original weights.The AION version.
You don't download the provider's original weights. You download models AION has quantized, adapted and tuned to run on the machine in your company.
Explore AION modelsQwen 3.8 Flash NextNVFP4AION
- Base
An open-source model
We start from models by the major providers, open and verifiable.
- Quantization
Lighter weights
Weights drop to 4 bits: even large models fit in the machine's memory.
- AION optimization
Ready for your machine
Fine-tuning where needed and inference engines tuned by AION.
- No usage limits
- Query the models as much as you need: usage is unlimited.
- Always up to date
- With the AION On-Premise licence the catalog stays up to date.
Power cut?It all comes back on its own.
AION Hub runs as a system service. When the machine powers back on, Hub restarts, brings back the models that were running, and then Agents is back in service.
Nobody has to walk into the server room or press Start again. Models you had stopped stay stopped.
- 01Machine back onUp
- 02AION HubUp
- 03Qwen 3.8 Flash NextRunning
- 04Qwen3 EmbeddingRunning
- 05Qwen3 RerankerRunning
- 06GLM-OCRRunning
- 07Whisper Large v3Running
- 08OpenAI-compatible gatewayUp
- 09AgentsUp
Let's see HubOn your machine.
We'll show you the catalog, starting the stack and the recovery after a restart, on a real case from your company.
- [email protected]
- Office
- Corso Vercelli 89, 28100 Novara