demi is not another AI model but the layer AI models work on. That is why demi ties itself to no single model: the AI gateway HALO picks a fitting one for every task.
One gateway, many models
Every model call demi makes goes through HALO. The gateway decides by three rules:
- EU by default. Models run through EU endpoints by default. Other regions only if you agree them with us in your contract.
- The cheapest model that does the task safely. Large frontier models only come in for hard tasks.
- Interchangeable. When a better model arrives, we swap it in behind the same interface. Nothing changes on your side.
Wired today
A selection of the models HALO can use today. The list grows as better models arrive.
- Language and reasoning
- Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5; GPT-5.2 and GPT-5 nano; open models such as Qwen3.5, gpt-oss-120b, Llama 3.3 70B and Mixtral 8x7B
- Code
- Qwen3-Coder
- Create images
- GPT Image, FLUX.2 pro, FLUX.1 Kontext pro, Stable Diffusion XL
- Understand images
- Qwen2.5-VL
- Search
- Embeddings with text-embedding-3-large, bge-m3 and Qwen3-Embedding; reranking with Cohere Rerank
- Speech to text
- Whisper large-v3-turbo
- Safety
- Qwen3Guard to classify content
The full catalogue with each model's strengths is under AI models.
Who picks the model?
- In the app and for agent runs through the API: HALO, automatically for every task. Your tenant's rules always apply.
- In the CLI: HALO as well, as long as you name none. With
-myou pick a model your tenant has allowed. - In an AI client over MCP: the client's model. Claude uses Claude, ChatGPT uses an OpenAI model. demi supplies the tools and the data.
Terminal
demi ask "Summarise this week's complaints." # model: autodemi ask -m <alias> "Draft the reply to the customer."Your data
- demi trains no AI models on your data, unless you explicitly want it to.
- Which providers work from which locations is in the sub-processor list. The Security page gives the overview.
