Every model on NVIDIA NIM, and which ones are free
The short answer
NVIDIA's hosted NIM endpoint serves 102 models from 25 vendors. Only 44 are NVIDIA's own. The rest are Meta, Google, Mistral, DeepSeek, Microsoft, IBM, OpenAI and more, all behind one OpenAI-compatible URL.
Read the list yourself at any time:
curl https://integrate.api.nvidia.com/v1/models \
-H "Authorization: Bearer $NVIDIA_API_KEY"Free? Free to try on NVIDIA's credit pool, paid after that. A pooled free key keeps them reachable with no card.
The most common misconception about NVIDIA NIM is that it serves NVIDIA models. It does, but they are a minority of the catalog. NIM is a packaging and serving layer, so what you actually get is a single OpenAI-compatible endpoint in front of most of the notable open-weight model families at once. That makes it one of the better places to compare models without changing any code between them. Here is what is actually in there.
The catalog by vendor
Counted from a live call to the models endpoint:
| Vendor | Models | Worth knowing |
|---|---|---|
| NVIDIA | 44 | Nemotron in every size, plus guard and embedding models |
| Meta | 10 | llama-3.3-70b-instruct, the dependable default |
| 9 | Gemma family, including gemma-4-31b-it | |
| Mistral | 7 | Codestral for code, Mixtral, Mistral Large |
| IBM | 4 | Granite, including two code-specific models |
| Writer | 4 | Palmyra, domain-tuned for finance and medicine |
| DeepSeek | 3 | deepseek-v4-pro and deepseek-v4-flash |
| Microsoft | 3 | Phi 3 family, including a vision model |
| OpenAI | 2 | gpt-oss-120b and gpt-oss-20b, the open-weight releases |
| 16 others | 16 | Kimi, GLM, Granite, Yi, Jamba, Snowflake, StepFun and more |
The Nemotron family, decoded
There are 24 Nemotron variants in the catalog, and the naming is the confusing part. Nemotron is NVIDIA's own line, generally built by post-training an open base model (often Llama) for reasoning and instruction following. The size words in the name are what matter:
| Tier | Example id | Use it for |
|---|---|---|
| Mini / Nano | nvidia/nemotron-mini-4b-instruct | Latency-sensitive work, edge deployment |
| Nano (MoE) | nvidia/nemotron-3-nano-30b-a3b | 30B total but only 3B active, so cheap to run |
| Super | nvidia/nemotron-3-super-120b-a12b | The general-purpose reasoning tier |
| Ultra | nvidia/nemotron-3-ultra-550b-a55b | Hardest reasoning, slowest, most expensive |
| Guard / Safety | nvidia/nemotron-3.5-content-safety | Classifying content, not chatting |
The a in a name like 120b-a12b means active parameters: it is a mixture-of-experts model with 120B total weights but roughly 12B active per token. That distinction decides your hardware, not the headline number, which is the subject of a separate post on why parameter count predicts VRAM badly.
A trap worth naming: several Nemotron entries are not chat models at all. nemotron-3-embed-1b and llama-nemotron-embed-1b-v2 are embedding models, nemotron-4-340b-reward is a reward model, and nemotron-parse is a document parser. Sending them a chat completion will not do what you expect.
Which ones are actually free?
All of them, briefly. None of them, permanently. NVIDIA gives new accounts on build.nvidia.com a pool of free inference credits that works across the whole catalog, with no card required. That is an evaluation allowance rather than a free tier, and once it is spent the hosted endpoint needs billing. Self-hosting a NIM container is the other route, and it needs an NVIDIA AI Enterprise license plus a GPU large enough for the model.
If you want NIM models to stay reachable at zero cost, the practical answer is to call them through a gateway that pools free tiers across several providers. The request body does not change, only the base URL and the key:
curl https://dreamprompting.com/api/v1/chat/completions \
-H "Authorization: Bearer YOUR_FREE_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/meta/llama-3.3-70b-instruct",
"messages": [{"role": "user", "content": "Hello from NIM."}]
}'The nvidia/ prefix pins NIM. Drop it and send "model": "auto" to let the gateway choose a healthy provider instead, which matters more than it sounds: NIM's larger models cold start, and failover turns a multi-second stall into a served request. Getting a key takes a few seconds on the account page, with no card.
Picking a model, briefly
- General chat:
meta/llama-3.3-70b-instruct. Reliable, well understood, widely benchmarked. - Reasoning and multi-step work: a Nemotron super or ultra tier, or
deepseek-ai/deepseek-v4-pro. - Low latency:
meta/llama-3.1-8b-instruct. Small models stay warm, so they dodge cold starts. - Code:
mistralai/codestral-22b-instruct-v0.1or the IBM Granite code models. - Vision:
meta/llama-3.2-90b-vision-instructormicrosoft/phi-3-vision-128k-instruct. - Open weights you intend to self-host later: check it fits first with the VRAM calculator.
Read the list live, not from here
NVIDIA adds and retires microservices continuously, so any published list including this one starts drifting immediately. Treat the numbers above as a snapshot taken on 2 August 2026 and read the catalog fresh before you hardcode an id. One call does it, against either endpoint:
# NVIDIA's own catalog
curl https://integrate.api.nvidia.com/v1/models -H "Authorization: Bearer $NVIDIA_API_KEY"
# What this gateway can currently route to
curl https://dreamprompting.com/api/v1/models -H "Authorization: Bearer YOUR_FREE_KEY"New to NIM entirely? The getting-started tutorial walks from account creation to a streaming call, and the endpoint reference covers the base URL and the speed tradeoff in more depth. You can also browse the full 279-model index across every lab, not just what NIM serves.