Train adapters
What to expect
This page lists, in one place, the limits you can meet when Mubit trains adapters for your calls on app.mubit.ai today. It is for anyone planning on adapters, or wondering why one doesn't do more.
Which calls adapters answer
| Limit | Today |
|---|---|
| Calls an adapter can answer | Chat Completions calls on /openai/v1 that go out on an OpenAI key or a Neon AI Gateway key. Calls to Anthropic, Gemini, the OpenAI Responses API and the OpenAI-compatible services go to the provider. See Calls answered by an adapter. |
| Size of a call | At most 256 KiB, and within the adapter's context of 8,192 tokens. |
| Content | Text only: a call with an image or a document goes to the provider. |
| Recording | Only calls Mubit records in full. |
| Availability | Adapters run on capacity Mubit shares across the workspace. When it isn't ready, your provider answers. |
| Routing decisions | The workspace has a daily limit on routing decisions. Past it, calls go to your provider until the next day. |
| Extra wait | A call that could go to an adapter can wait up to about a second for the decision. |
Training
| Limit | Today |
|---|---|
| Base model | Qwen/Qwen3-4B-Instruct-2507, for every adapter. |
| First training | At 25 kept examples, with at least 12 more kept back for the exam. |
| New version | After at least 25 new kept examples, at least half as many as the live version trained on, and at least a day after the last training. |
| Examples per training | At most 250 kept examples, spread across conversations, newest first. A kept example over 256 KiB is left out, and so is one whose reply states a figure its request doesn't hold. |
| Training capacity | Shared by every organization in the workspace: a few trainings at a time, and a daily budget that resets at 00:00 UTC. |
| How long it takes | It depends on the queue. The task's page shows each stage as it happens. |
See Training.
Rating
| Limit | Today |
|---|---|
| Automatic rating | Up to 200 calls of each task a day. |
| Your ratings | up or down, with an optional comment of up to 500 characters. The latest rating counts. |
See Tasks and ratings.
The exam and the rollout
- The exam advises. A version below its target can still go into the rollout, and its replies show
below its testin Requests. See How the exam acts today. - Mubit moves a version through the rollout on counts: 50 compared replies from 10 conversations for the silent test, then 100 calls in the partial rollout.
- Mubit doesn't take a version offline or roll back on its own. You do this in the console.
See The rollout.
What you can do where
| Action | Console | SDKs, CLI, management API |
|---|---|---|
| Rate a call | Yes | Yes |
| Start a silent test or a partial rollout, change its share | Yes | Yes |
| Make a version live | Yes | Yes |
| Roll back | Yes | Yes |
| Take a version offline | Yes | No |
| Switch an adapter off or on | Yes | No |
| Emergency stop | Yes | No |
The management API answers 501 for the operations marked No, and for pausing or resuming a task's training. To start a training now, use the SDKs, the CLI or the management API: see Start a training yourself.
Try models
Try models runs provider models on Mubit's key, and your adapters beside them, with a number of calls for your organization each day. Its calls are not stored and train nothing. See Create your first adapter.