Test-Time Training via one tool-call
One call to internalize() updates the model weights so it deeply knows and understands your knowledge with zero context.
GLM 5.3 available now. More models coming.
How learning works1 / 4
Teach the model something it needs to know
Pass the knowledge your agent needs to internalize(). For example, a support agent could learn that your company accepts returns within 21 days.
How learning works2 / 4
The knowledge is learned through an adapter
We use test-time training to encode the policy in adapter weights, so the model can reason with what it has learned when answering questions.
How learning works3 / 4
The next conversation doesn’t need the document
When you start a fresh conversation using that memory, the model has the policy available through its weights. You don’t have to add the original text to the prompt.
How learning works4 / 4
The model can apply the policy to a new situation
If a customer asks about an order placed 17 days ago, the model can use the 21-day policy to work out that they still have four days to return it.
How serving works1 / 4
Your agent uses the memory that’s ready to serve
Requests for a subject use its active adapter. You keep calling the same API while we host the model and the memory it needs.
How serving works2 / 4
New knowledge can be learned while the agent is running
Another internalize() call creates a new adapter version for that subject. The current version continues answering requests while the new one trains.
How serving works3 / 4
We route requests to the new version when it’s ready
Once the new adapter is ready, new requests for that subject use it automatically. Your application keeps the same subject ID and API calls.
How serving works4 / 4
Internalize handles training and serving for you
We manage the adapter versions, host them, and handle the switch between them. Training and inference run serverlessly, so there’s no model server for you to deploy or maintain.