AI model providers
The providers Remy relies on, the models each one runs, and what happens to your data at each.
Last updated: August 28, 2026
The rule we hold providers to
No training on your data — at any provider, ever. Every provider that processes your conversations, mail, or calendar is contractually barred from training models on your data. Most of what Remy does runs under zero-data-retention terms: nothing is kept after your request completes. Where a lane runs on other terms — an enterprise arrangement, an exception request still pending — this page prints the exact state rather than rounding up. That includes the requests we have filed and not yet had granted.
Text and understanding
Remy’s primary inference provider. Runs the assistant you talk to (DeepSeek, with GLM as its sibling seat), the background readers that index and understand your mail and calendar (DeepSeek), image and document reading (Qwen), and search embeddings. No training on your data; nothing retained after the request. Read Fireworks’ data handling policy.
Outage backup running the same models, so a provider outage degrades to the same brain on different servers — not a different assistant. No-training terms. Used only when Fireworks is unavailable. Read DeepInfra’s data privacy policy.
The web-errand browser lane is moving to Gemini (decision August 22, 2026; it replaces the previous model on this lane). Terms: Google’s paid enterprise API — no training on your data, your data not mixed with anyone else’s. Google’s abuse monitoring can normally involve prompt logging and human review; we requested the exception to switch that off for Remy’s project, and Google granted it on August 28, 2026. So: no training, no co-mingling, and no prompt logging for abuse review. (We filed this before the lane went live and printed “pending” here until the approval came - that is how this page works.) Read Google’s enterprise data-governance terms.
Voice
Turns the mic button’s dictation into text (Whisper). Zero-data-retention account. Read Groq’s data policy.
Live voice conversations: LiveKit carries the session and transcribes your speech; Cartesia speaks Remy’s replies. Both are configured not to train on your data; we are completing verification of their retention terms and will update this page with the result.
When Remy places a phone call for you, the in-call agent runs on Retell’s stack, which uses an OpenAI model hosted by Retell and ElevenLabs voices. This lane predates our zero-retention standard and is under active review; calls are always placed with your explicit approval, and no mail or calendar content is shared with the call stack beyond what the call itself requires.
Scope
This page covers AI model providers. The full vendor list, including hosting and infrastructure, lives on the subprocessors page. Questions: [email protected].