Nebul Docs

Model Lifecycle

Lifecycle stages for models on the Nebul Inference API, how model version changes and deprecations are handled, and how to check them from the API.

Every model on the Nebul Inference API sits in one of three lifecycle stages. This page explains what each stage means, how we introduce new model versions, how deprecations work, and how to check the current state from the API.

Lifecycle Stages

Production

Production models are ready for real workloads and run on dedicated GPU infrastructure. They've been through performance validation, so you can build on them with confidence.

  • Stable API surface and model behavior
  • Full monitoring and alerting
  • No breaking changes without advance notice
  • Recommended for production use

Preview

Preview models are open to all customers while validation is still underway. They give you early access to new capabilities, with a few trade-offs while we finish testing. Preview models are not suitable for production use.

  • Model weights or configurations may be updated any time
  • Expect unavailability as preview models are regularly updated with configuration or serving changes
  • Will be promoted to Production or deprecated once validation is complete

Early Access

Early Access models are shared with select customers on request, ahead of general availability. They typically run on experimental infrastructure while we actively tune them, and we make them available so you can evaluate and give feedback early.

  • Access is granted per organization
  • Runs on experimental or non-production infrastructure
  • May be interrupted, restarted, or reconfigured without notice
  • Model weights, quantization, and serving parameters can change frequently
  • Not recommended for production use
  • May be deprecated without a migration path

Important

Early Access models are meant for evaluation and experimentation only. Do not use them for production workloads.

How to Tell Which Stage a Model Is In

Check the Model Catalog and AI Studio Inference → Models. Preview models show a Preview badge on the card; Studio also has a Show preview models filter. Production models have no lifecycle badge. Early Access models are not listed in the public catalog; we tell you directly when access is first granted.

How Model Changes Are Handled

Changes to the model catalog fall into three categories. Each has a defined process and notice period, so you can plan around them.

In-place updates

An in-place update changes the weights or serving configuration behind an existing model ID, for example when a vendor ships a patch release. The model's behavior contract stays the same. In-place updates are announced in the Changelog if relevant.

The model_version field in /v1/model/info pins the deployed revision (a HuggingFace repo sha) when set; a change to it means the model you're calling was updated in place.

New model versions

A new flagship version ships as a new model ID, for example zai-org/GLM-5-FP8 followed by zai-org/GLM-5.3. We never swap a new version silently into an existing model ID.

Before a new version goes live, we publish what changed:

  • Reasoning behavior: the default and available reasoning_effort values, how the model thinks, and which field returns the reasoning (see Reasoning)
  • Output characteristics: response style, verbosity, and formatting
  • Operating envelope: context window, token pricing, and latency profile
  • The overlap window: how long the previous version stays callable

During the overlap window both model IDs answer requests. The previous version keeps working until its published removal date, and its /v1/model/info entry carries deprecation_date and superseded_by_model_name from the day the new version is announced. Switch whenever you're ready within the window.

Reroutes

When a successor is a drop-in replacement, the same model lineage with compatible behavior and comparable pricing, we may route an existing model ID to the new version. Requests to the old ID then return responses from the new model. Reroutes are announced on this page and in the Changelog.

The three change types at a glance

In-place updateNew model versionReroute
Model IDUnchangedNew ID; the old ID keeps answering until its removal dateUnchanged
Signal in /v1/model/infomodel_version changesThe old entry gains deprecation_date and superseded_by_model_nameNone; the ID starts answering from the successor
Action neededNoneSwitch to the new ID before the removal dateNone required; verify outputs still fit your use case
Where it's announcedChangelogChangelog and this pageChangelog and this page

Deprecation Policy

Lifecycle stageNotice before removal
ProductionAt least 14 days
PreviewWe aim for the same 14 days, wherever possible
Early AccessMay be removed at any time

When a model is scheduled for deprecation:

  • The model stays callable until its removal date, so you have time to migrate
  • After the removal date, requests to the model return 404 with a model not found error
  • Its /v1/model/info entry carries deprecation_date and superseded_by_model_name
  • The deprecation is announced on this page and in the Changelog
  • Where a close equivalent exists, we typically reroute the deprecated model ID to it at a comparable price

We recommend switching to the successor as soon as a deprecation is announced.

What a deprecation looks like over time

  1. Announcement. The deprecation is published on this page and in the Changelog, and the model's /v1/model/info entry gains deprecation_date and superseded_by_model_name.
  2. Notice window. The model keeps serving exactly as before. Nothing about calls to it changes, and code that hasn't migrated keeps working until the removal date.
  3. Removal. On the removal date, requests to the model start returning 404:
{"detail":{"message":"model 'Qwen/Qwen3.5-27B': not found","error_type":"not_found"}}

If you hit this error, the model has been removed: check the deprecation history below for what replaced it, or ask about dedicated endpoints if you need that exact model back.

Scheduled Deprecations

Models currently scheduled for removal. Each stays callable until its removal date.

AnnouncedRemoval dateModelSuccessor

/v1/model/info is the source of truth for your project. Models enabled for your project that aren't in the public catalog, such as Early Access models, only appear in the keyed response.

Deprecation History

Models already removed from the Inference API, most recent first. A blank successor means no direct equivalent was designated; use the Model Catalog to find a current model for your workload.

RemovedModelSuccessor
August 30, 2026zai-org/GLM-5.2-FP8zai-org/GLM-5.3
June 29, 2026vidore/colqwen2-v1.0-merged
June 16, 2026zai-org/GLM-5.1-FP8-Nebulzai-org/GLM-5.2-FP8
June 16, 2026mindlab-research/Macaron-V1-Preview-749Bzai-org/GLM-5.2-FP8
June 16, 2026Qwen/Qwen3.5-27Bnvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8
June 10, 2026Salesforce/SFR-Embedding-Mistral
June 5, 2026mistralai/Voxtral-Mini-4B-Realtime-2602CohereLabs/cohere-transcribe-03-2026
May 5, 2026deepseek-ai/DeepSeek-V4-Promistralai/Mistral-Large-3-675B-Instruct-2512
May 1, 2026meta-llama/Llama-4-Maverick-17B-128E-Instructmistralai/Mistral-Large-3-675B-Instruct-2512
May 1, 2026Qwen/Qwen-Image-2512Tongyi-MAI/Z-Image-Turbo
May 1, 2026zai-org/GLM-ImageTongyi-MAI/Z-Image-Turbo

Check from the API

List every model available to your project that is scheduled for removal:

import os

import requests

response = requests.get(
    "https://api.inference.nebul.io/v1/model/info",
    headers={"Authorization": f"Bearer {os.environ['NEBUL_API_KEY']}"},
)

for model in response.json()["data"]:
    info = model["model_info"]
    if info.get("deprecation_date"):
        print(model["model_name"], "->", info["deprecation_date"],
              "| successor:", info.get("superseded_by_model_name"))

A model scheduled for removal lists its deprecation_date and superseded_by_model_name. An empty result means no model available to your project is currently scheduled for removal.

Run the check on a schedule and treat any result as actionable: it lists every model your project can call that has a removal date. Catching it early turns a deprecation into a planned migration instead of a production 404.

See Model Catalog for the full field reference.

Migrating to a Replacement

Keep model IDs in configuration rather than in code, so a migration is a config change instead of a release.

  1. Find the successor in the Model Catalog, or ask us for a recommendation through the Support Portal.
  2. Compare candidates on capability, latency, and cost for your workload. If the successor is a new model version, read the published differences first, the reasoning behavior in particular.
  3. Run an A/B test on representative traffic. Replacement models rarely behave identically, so validate output quality and token usage before committing.
  4. Shift traffic gradually: canary first, then partial, then full cutover. Keep a fallback path where you can.
  5. Watch error rates, latency, and cost during and after the migration.

If you need a model that is no longer in the shared catalog, contact your account team about dedicated endpoints.

Requesting Access

Want Early Access to a specific model, or think a Preview model is ready for Production? Reach out through the Support Portal or contact your account team. We're happy to talk it through.

On this page