Reading a 60+ Model Catalog Without Guessing
Prerequisite: Lesson 01's finding that this site's own existing Nebius lessons already use model names absent from Nebius's current catalog pages. After this lesson, you can: query Token Factory's own live model list and deprecation notices instead of trusting any snapshot, including this course's, and you'll have watched this course catch a real, dated problem in this site's own prior work while doing it.
The proof is this site's own code
Lesson 01 found that the model names in this site's existing Nebius lessons don't match Nebius's current marketing pages. That was a soft finding: names drift, pages disagree, nothing was dated or official. This lesson has a harder one.
meta-llama/Llama-3.3-70B-Instruct is the model lesson 08 of tools-memory-and-multi-agent-systems
uses in a shipped, lint-passed CLI-wrapper lab, dispatched through real working code against
api.tokenfactory.nebius.com. Nebius's own August 2026 deprecation notice, fetched directly from
docs.tokenfactory.nebius.com while writing this lesson, lists that exact model ID as discontinued
on August 31, 2026, with a named replacement: Qwen/Qwen3-30B-A3B-Instruct-2507. This isn't a
guess about catalog churn. It's a dated fact, published by the vendor, about a piece of code this
site has already shipped.
The pattern, not the incident
One deprecated model would be a curiosity. Twelve models retiring on the same date, all with named
replacements, is a pattern: Nebius runs deprecation as a routine, scheduled process, not an
occasional surprise. The same August 2026 notice also lists MiniMaxAI/MiniMax-M2.5,
NousResearch/Hermes-4-70B, three separate nvidia/Nemotron variants, Qwen/Qwen2.5-VL-72B- Instruct, Qwen/Qwen3-32B, Qwen/Qwen3-Next-80B-A3B-Thinking, and deepseek-ai/DeepSeek-V4-Flash,
each with its own named successor. A separate June 2026 notice retired eleven more models, effective
that same month, on a completely different set of names.
The June notice is the sharper warning of the two: it lists no replacement at all for any of its eleven models, and states outright that nothing reroutes automatically. A model going away with no named successor and no automatic migration is the worst case this catalog produces, and it already happened once, months before this lesson was written.
Querying the catalog instead of trusting a page
Every model page, every course (including this one), and every blog post is a snapshot the moment it's written. The one thing that stays current is the API itself.
curl -H "Authorization: Bearer YOUR_TOKEN" \
https://api.tokenfactory.nebius.com/v1/modelsReturns a ListModelResponse: an array of model objects, each with id, created, owned_by, and
an optional status field. This is the actual live catalog, not a page someone wrote and forgot to
update.
A model can still appear in GET /v1/models while carrying a scheduled, dated retirement. The
list-models response alone doesn't surface that. Check docs.tokenfactory.nebius.com's deprecation
notices directly (they're dated and versioned by month, as this lesson's own two receipts show)
before committing to a model for anything long-lived.
Every model string in every lesson of this course was correct on the date this lesson was checked. Whether it still is by the time you're reading it is exactly the question the two steps above answer, and no course can answer it for you permanently.
python list_live_models.py
Why this course caught its own drift, and what that's worth
This lesson exists because this course's own authoring process ran the check it's now teaching. That check found this site's own prior lesson using a model with a dated, named expiration, months before that model actually breaks anything. That's the entire value of the habit: catching a stale dependency while it's still a calendar entry, not an outage. A course that can name its own site's real gap, with a real date attached, is doing more for a reader than one that only warns in the abstract that "things change."
This is the last lesson in the current map. Check the course page for what's live and what's coming next.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
