Docs/Llmops/Model Versioning

Model Versioning & Deployment

ANTS Platform does not host or deploy models, and it has no model registry, deployment pipeline, blue-green or canary deployment API. Your provider or your own infrastructure serves the models. This page shows how to version and roll out model and prompt changes with the features ANTS Platform does have.

ℹ

For the model catalog with lifecycle status and end-of-life dates, see Model Lifecycle.

Version prompts and their model settings

Prompts are versioned. Each version can carry a config object, for example the model name and temperature. Your application fetches the version labelled production and reads the model from its config, so changing the model is a new prompt version plus a label move. Roll back by moving the label back.

python

from ants_platform import get_client ants = get_client() prompt = ants.get_prompt("support/answer", type="chat") model = prompt.config.get("model", "gpt-4o-mini")

Route between models with Budget AI

With Budget AI, applications call a model group name such as prod-chat instead of a provider model. Behind the group you choose the provider models:

Switch models without a code change. Change the deployments in the group.
Gradual rollout. Add the new model to the group with a small weight and raise it over time, using the weighted routing strategy.
Fallbacks. Send requests to a fallback group when the primary models fail.

See Model Groups & Routing.

Tag traces with your application version

Record which version of your application produced each trace, so you can compare before and after a change:

Pass version= on observations or in update_current_trace(version=...).
Set release= on the client, or the ANTS_PLATFORM_RELEASE environment variable, for the deployed release.
Set environment= on the client, or ANTS_PLATFORM_TRACING_ENVIRONMENT, to separate staging from production.

python

from ants_platform import AntsPlatform ants = AntsPlatform(release="2026.10.2", environment="production") with ants.start_as_current_span(name="answer", agent_name="support-agent", version="v2"): ...

The tracing tables can filter by version, release and environment. Dashboards can filter by environment.

Compare before you switch

Run an experiment on a dataset with the current and the candidate model, scored by LLM-as-a-Judge.
Use the AI Optimizer to test comparison models against an agent's real traces.
After the switch, watch cost and latency per model on the Cost Dashboard and Latency Dashboard.

Next steps

Planned capabilities are listed on the roadmap.