Getting started

helexa serves open-weight models on hardware we own and operate in the EU. You can use it two ways: through the web app, or through the API with any OpenAI- or Anthropic-compatible client.

In the browser

Nothing is required to start a conversation — open the chat and type. Conversations are stored in your browser, not on a server, so they survive reloads on the same device and go no further.

Creating an account adds an API key, usage visibility, and higher limits. It does not move your conversations anywhere.

From code

Point any OpenAI-compatible client at https://helexa.ai/v1 and use an API key from your account.

curl https://helexa.ai/v1/chat/completions \
  -H "Authorization: Bearer $HELEXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "helexa/balanced",
    "messages": [{"role": "user", "content": "Say hello."}]
  }'

The same key works against the Anthropic-shaped /v1/messages endpoint and the OpenAI /v1/responses endpoint. See the API reference for the full surface.

Choosing a model

Ask for a tier rather than a specific model, and you get whatever currently fills that role:

Tier	For
`helexa/small`	short, cheap, fast turns
`helexa/balanced`	general use — the sensible default
`helexa/large`	long context, harder reasoning, agentic work
`helexa/image`	text-to-image

Tiers are stable; the model behind a tier is not. That is the point — you get upgrades without changing code. When you need a specific model and want it to stay that way, use its full id from GET /v1/models.

Limits

Public API requests are rate-limited to 10 per minute per IP, with a short burst allowance. That suits interactive use and normal application traffic. If you are running a batch job or a busy service against helexa, talk to us rather than working around the limit.

Models load on demand. A request for a model that is not currently resident on a GPU may take tens of seconds while it loads, then run at full speed. Clients should allow a generous timeout — 300 seconds is what our own tooling uses.

What runs where

Every request is served on hardware helexa operates. Nothing is forwarded to a third-party inference provider, and prompts are not used for training. See privacy for the specifics.