Getting started

1. Create an account

Go to sovrgpt.com/login.
Enter your email → you receive a magic link (no password, no credit card).
Open the link — done. A personal workspace is created automatically.

2. Start your first chat

Pick a model at the top right (the default is gemma-4-12b with a 262k context).
Type your question and, if you like, enable tools in the composer:
- Web search (Brave — 🔴 US-based provider, opt-in per chat, off by default; see Data protection) — the same switch also lets the model open and read a page (fetch_url), for example a URL you pasted into the chat.
- Image generation (Z-Image / FLUX.2)
- Video generation (LTX-Video)
- Marketplace connectors (GitHub, Notion, …)
The answer streams back. A cold start only happens when the model has not been used for a while — on the default tier that has been measured at 5.5 to 9.5 minutes (three runs: 332 s / 504 s / 574 s), and similarly on the larger tiers. When warm, the model answers in under 10 seconds.

3. Switching models

Saved per chat. The tiers running on our own hardware:

Tier	Model ID	Context	Usage profile
`default`	`gemma-4-12b`	262k	General purpose, understands images
`balanced`	`qwen3.8-27b`	131k	More depth, strong at coding
`vision`	`gemma-4-26b-a4b`	262k	Understands images
`coder-mini`	`qwen3.6-35b-a3b`	131k	Agentic coding through IDE extensions
`reasoning`	`qwen3.5-9b-deepseek-v4-flash`	16k	Visible chain of thought
`llama`	`llama-3.1-8b`	16k	Llama-specific workloads. ⚠️ No tools / web search

🔑 The strongest model on our own hardware is qwen3.8-27b — not the one with the most impressive tier name. Tier keys are technical identifiers from the early days, not a quality promise.

🔴 Withdrawn on 2026-09-11: qwen3.5-35b-a3b (premium), qwen3-coder-next-fp8 (coder) and qwen3.5-9b (default-legacy). Both had been overtaken by active models — premium had the smallest context window in the catalog at 8,192 tokens, and coder lost to qwen3.6-35b-a3b at a quarter of the context. Existing integrations do not break: the role names coder, coding, qwen-coder and qwen-premium now point at the successors, and anyone sending the exact old model id gets the default tier plus the x-sovrgpt-model-substituted header (see Chat Completions).

⚡ If you need an immediate answer: every tier above scales to zero when idle — the first call after a pause waits 3 to 15 minutes for the cold start. The models in the European partner data centre (tensorx-…, stackit-…) have no cold start. The full list with measured cold-start figures is in Models.

4. Create an API key

For your own integrations (scripts, apps, automations):

Settings → API keys.
"New API key" → give it a name.
The key is shown in clear text once — copy it immediately.
Org-scoped: the key works in the context of the active org (quota, logs, RLS).

5. Your first API call

curl https://sovrgpt.com/api/v1/chat/completions \
  -H "Authorization: Bearer $SOVR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-12b",
    "messages": [{ "role": "user", "content": "Hello SovrGPT!" }]
  }'

Full API documentation: API reference.

6. Invite a team org (optional)

If several people should work on the same chats, connectors and API keys:

Click "+ New team organisation" in the sidebar.
Choose a slug and a name.
Invite members by email and assign roles (owner / admin / member).

Common pitfalls

Cold start: the first call to a model that has been idle for a while takes 30 s to 5 min for the first answer (depending on the tier). Follow-up calls are warm.
Lost API key: it cannot be displayed again — create a new key and revoke the old one.
403 in the marketplace: the connector probably requires an admin role in the team org.