Technology
AI models & BYOK
One request, one provider — chosen by task, with failover only before the first token, and keys that never leave the server.
Routing live in buildsBYOK planned
- Exactly one provider per request. The next in the chain runs only on temporary or capacity failure — and only before the first token streams.
- Model ids live in one registry. This lineup can change — when it does, this page changes with it.
The lineup
What runs where today
| Capability | Primary | Fallback |
|---|---|---|
| Text & reasoning | Groq — openai/gpt-oss-20b | OpenRouter — nvidia/nemotron-3-ultra-550b-a55b:free |
| Vision (screen answers) | Groq — qwen/qwen3.8-27b | Cloudflare Workers AI — @cf/google/gemma-4-26b-a4b-it |
| Grounding (target boxes) | Google Gemini — gemini-3.8-flash | Cloudflare Workers AI — @cf/google/gemma-4-26b-a4b-it |
| Speech to text | Groq — whisper-large-v3-turbo | — |
| Text to speech | Cloudflare — deepgram/aura-1 | Groq — canopylabs/orpheus-v1-english |
Model ids live in exactly one registry in the app, and this lineup can change — when it does, this page changes with it. Nothing here is a guarantee of future availability.
Failover rules
How failover behaves
-
Pre-token only
Mid-stream, the provider is committed; you never get two models stitched into one answer. Aborts are never failovers — pressing Stop stays a stop.
-
Cooldowns
Cooled-down providers are skipped for 60 seconds (rate/quota) or 30 seconds (transient), so a dead chain routes straight to the healthy one.
-
Terminal stays terminal
Auth and validation failures never loop through providers. Failures surface as short, sanitized notes — never as fake successes.
BYOK — bring your own key
Status: planned, not implemented. The router registry already reserves a slot for future BYOK providers, and the design rule is written down: when BYOK ships, keys go into the operating system’s credential storage — never into the chat database, never into a log, never into the renderer’s local storage. Until that exists, DOT404 routes through its own gateway, where provider keys live in server-side function secrets only.
No BYOK UI, provider list, or pricing exists today. See the roadmap for how it is sequenced.
Signed out = local preview
Without a session there is no gateway, so replies come from a local mock provider that is clearly labeled as such — word-by-word streaming over your topic, no network, no cost. It exists so the app stays honest and usable offline, not so anyone can mistake it for a real answer.
Bring DOT mini home.
Free for Windows 10 and 11. One installer, two bots, zero clutter.