Skip to content
Period 1 / 6

Week 0: The Model Name You Just Copied Is Already Wrong

Before you start

Prerequisite: none. This is a five-minute orientation, not a lesson. After this, you can: decide whether this course is worth five lessons of your time, before spending any of them.

A model ID this course itself got wrong once

Lesson 03 originally shipped naming nvidia/llama-3.1-nemotron-70b-instruct as its reasoning-model example. tutorial-courses-nim-review, an independent adversarial currency review run the same day this course published (the process RULING-069 requires before either NVIDIA course ships), found that exact ID already deprecated and dropped from NVIDIA's own live API reference. The tutorial this course replaces named llama3-8b-instruct back in 2024, an ID that's now two catalog generations stale.

That's not a knock on the source material. It's the actual shape of NVIDIA's NIM (NVIDIA Inference Microservice) catalog: models get added, renamed, and retired often enough that a code sample frozen at publish day starts going wrong within months, sometimes the same day. If you've got a meta/llama-... or nvidia/... model string copied into a script right now, there's a real chance it's already stale.

What doesn't move, even while the catalog does

Here's the part worth knowing before lesson 02: NIM's chat, embedding, reranking, and generation models all sit behind the same OpenAI-compatible client shape. One OpenAI client object, one base_url, one auth header. The model ID is the only thing that changes between a small instruct model and NVIDIA's largest reasoning model. Learn that shape once and catalog churn stops being a rewrite. It becomes a config value you update.

What churns, and what the client pattern absorbs

This course teaches you that pattern directly, then shows you where to check a model's current status yourself, so you're not depending on this course (or any course) to stay current for you forever.

Quick check — NVIDIA retires the reasoning model your NIM client currently calls and replaces it with a newer one. Under the client pattern this course teaches, what has to change in your code?

Why lesson numbering starts at 02, not 01

One housekeeping note before you start: this course's lessons are numbered 02 through 06. That's not a missing lesson 01, it's this Week 0 orientation taking that slot in spirit while keeping the original lesson IDs stable. Nothing was renumbered to make room for it.

Where this course is honest about a gap

Lessons 05 and 06 cover self-hosting a NIM container and sizing GPUs for production. Neither one invents a cost figure or a war story this course hasn't actually lived. No first-party receipt exists yet for running NIM in production at scale here, so those two lessons teach the current, correct, vendor-documented pattern and say so plainly, rather than fabricate a number to sound more finished than the material actually is.

What you can do after Week 1

After lesson 02 you can write a chat-completion client against any NIM-hosted model and swap the underlying model with a one-line change. By lesson 06 you can also size a GPU deployment against NVIDIA's own current guidance, and, more durably, check any model ID's live deprecation status yourself before you build on it, the same habit that caught this course's own stale example before it shipped.

Start Lesson 1

Chat Completion and the OpenAI-Compatible Client Pattern: the one client shape that survives every NIM catalog change.

Have a question about this lesson?

Reply here and it goes straight to Rod. Same as replying to one of his emails.