Bring Your Own Model

Bring Your Own Model#

By default, every Gydr widget runs on Gydr-hosted models. Bring Your Own Model (BYOM) lets you point a widget at a model on your own provider account instead — so inference is billed directly by your provider and your data stays within your provider relationship.

BYOM is opt-in per widget and available on the Growth plan and above. Auxiliary work — intent planning, history compression, language detection, and the async quality check — always runs on Gydr's infrastructure and is not affected by this setting.

Supported providers#

Nine providers are supported. Each one is available in the Console when you add a connection.

The model IDs shown in the Console are suggestions. You can type any model ID your provider supports — Gydr tests it live when you import it (see below).

How to configure#

Configuration happens in two places: first you add a connection and import models, then you attach a model to a widget.

Step 1 — Add a connection#

Open Console → Settings → Import Models. Click Add connection, choose your provider, and paste your API key. The key is encrypted immediately and only the last four characters are ever shown again.

Step 2 — Import a model#

Under the connection you just created, click Add model and enter the model ID. Gydr validates the model immediately by calling it with a test tool-call probe. A model that cannot be reached or does not emit a tool call stays in a validation failed state and cannot be selected in a widget. You can retry validation any time from the model row.

Step 3 — Attach to a widget#

Open the widget editor and go to the LLM Model tab. Active imported models appear under Your Models below the standard Gydr models. Select one to enable BYOM for that widget. Each widget has its own model selection — other widgets are unaffected.

Failure behavior#

Failure behavior controls what happens when your provider returns an error. You can set it per widget from the LLM Model tab.

BYOM entitlement is checked at chat time, not just at configuration time. If an account moves to a plan without BYOM access, widgets automatically fall back to their Gydr model. The widget continues answering — it does not go offline.

For how running your own model changes metering, see How Credits Work.

Things worth knowing#

• The model catalog is a suggestion list, not a whitelist. Any model ID supported by your provider can be imported. The live validation probe is what determines whether a model can be activated — not the list shown in the Console.
• Reasoning models work out of the box. Several providers (DeepSeek, Qwen, Moonshot Kimi thinking variants, OpenRouter, Groq) include models that emit extended reasoning tokens. Gydr handles the reasoning round-trip automatically — no configuration change is needed on your side.
• Moonshot Kimi thinking models require a large output budget. Models such as kimi-k2-thinking and kimi-k2-thinking-turbo need at least 16,000 output tokens to avoid truncating mid-reasoning. Gydr sets this automatically.
• Grounding and safety checks still apply. Your imported model goes through the same answer grounding gate as Gydr-hosted models. Responses that do not pass grounding are treated the same way regardless of which model produced them.
• Your API key is encrypted at rest. Keys are stored with AES-256-GCM encryption. The Console shows only the last four characters after initial entry. If you rotate the key on your provider's side, update it in Console → Settings → Import Models.