NVIDIA Gives Away 80+ Models for Free. I Read the Fine Print.

NVIDIA gives away access to more than 80 AI models through one OpenAI compatible API: DeepSeek, Kimi, GLM, and dozens more, with no GPUs to manage and no credit card, aimed squarely at people prototyping with tools like Cursor and agents.

The claim sounds like marketing, so I checked it against the live API. Most of it holds up. Here is what is actually in the catalog, what the fine print says, and the catch nobody puts in the headline.

Short version: the endpoint is real, the catalog today lists 81 models, and the price really is zero dollars. You pay in rate limits and in prompts. Read on before you point your agent at it.

What I actually verified

You do not need an account to check the catalog. The model list is public:

curl https://integrate.api.nvidia.com/v1/models

That returned 200 OK and 81 models, counted today, October 20, 2026. The headliners are in there:

• deepseek-ai/deepseek-v4.1-flash (DeepSeek’s current fast model, not the old R1 from the screenshots)
• moonshotai/kimi-k2.6
• z-ai/glm-5.3 and z-ai/glm-5.3-flash (GLM moved under Z.ai, the renamed Zhipu)
• openai/gpt-oss-20b, plus 17 Nemotron variants and the usual Llama, Mistral and Gemma back catalog

The setup in the post also checks out. Base URL is https://integrate.api.nvidia.com/v1, the key is a free NVIDIA Developer account key, and any OpenAI client works by swapping those two values. Cursor, Windsurf, your agent framework of choice: if it accepts a custom OpenAI endpoint, it accepts this one.

The catalog is a revolving door

MiniMax is not in the catalog today. Neither is Qwen, which was there earlier this year. NVIDIA rotates models in and out based on demand and licensing, and the slugs change when versions bump.

Practical consequence: pin nothing. If you build a prototype against moonshotai/kimi-k2.6, check the live list before you demo it, because the model you tuned against last month may have rotated out. The catalog is a buffet, not a contract.

What “free” means in October 2026

The old system, 1,000 inference credits at signup, is gone or vestigial depending on which NVIDIA page you read. What actually limits you now is rate: roughly 40 requests per minute per key, with a reported ceiling around 10,000 requests a day. There is no credit card. There is a phone verification step when you create the key, and it does not cover every country, which is the one real onboarding catch people hit.

Do the math on 40 requests a minute. That is 57,600 requests a day if you never sleep. For prototyping, evaluating a model on your own tasks, or running a personal agent, that is effectively unlimited. For a production chatbot it is a rounding error, and NVIDIA knows it: production points you at paid partner endpoints or self hosted NIM. This tier is a funnel, and the funnel ends at NVIDIA hardware and an AI Enterprise license. Nothing wrong with that, as long as you know which end of the funnel you are on.

The actual catch: your prompts are the product

Here is the part that gets buried under the caveats. On the free hosted endpoints, NVIDIA logs your inputs and outputs and can use them to improve its models. NVIDIA’s own guidance for the free tier is blunt: do not send confidential information or personal data.

So the rule I would run, and it is the same rule from my blast radius audit: a free endpoint is still inside your agent’s blast radius. Pointing Cursor at Kimi for free means your code transits NVIDIA’s servers and sits in their logs. Fine for open source, fine for throwaway prototypes, fine for the agent test loops where you are burning calls to see if a model can drive your tools. The wrong place entirely for your private repo, your client work, or anything with a secret in it. The self hosted NIM path is where your data stays yours, and of course that is the paid path. The free tier’s privacy policy is the price tag; it is just denominated in prompts instead of dollars.

Who this is actually for

Used right, this is the best model shopping tool going. Last month I ran DeepSeek and Qwen locally on a $5 VPS to see if the Chinese models are winning. This endpoint is the lazy sequel: try DeepSeek V4.1 Flash, Kimi K2.6 and GLM-5.3 on your real prompts first, for zero dollars, and only download weights for the one that earns it. It is also a genuinely good deal for agent development, where half your calls are the model fumbling tool syntax and you would rather not pay frontier prices for the fumbles.

Just do not mistake it for infrastructure. The catalog rotates, the rate limit is a hard ceiling, and your prompts are training data.

My three picks

Eighty one models is a catalog, not a shortlist. If you create a key, these are the three I would spend my first calls on:

• DeepSeek V4.1 Flash (deepseek-ai/deepseek-v4.1-flash): the free frontier coder. It is the current fast DeepSeek, with a million token context, and it is the natural sequel to the little R1 distill I ran on the $5 VPS: same family, wildly more capable, zero dollars. For code generation and reasoning on a budget, it is the first slug I would paste into Cursor.
• Kimi K2.6 (moonshotai/kimi-k2.6): the agent’s model. Moonshot builds Kimi for long, tool-using sessions, which is exactly the workload that eats API budgets: dozens of calls, half of them the model fumbling tool syntax. Fumbling is free here. If you are testing whether an open model can drive your agent loop, this is the test.
• GLM-5.3 (z-ai/glm-5.3): the all-rounder. On Z.ai’s own benchmark tables it sits in the same league as the closed frontier models, and while maker tables deserve salt, the free tier exists precisely so you can check the claim on your own prompts instead of arguing about it. If you only try one model from the catalog, try this one.

Short version: it is real, with footnotes. Eighty one models, one endpoint, no card, and a privacy policy doing the charging. For prototyping that trade is fine. NVIDIA is betting you will think so too, right up until you need the paid tier.