Nvidia's free AI model now handles more than 3 trillion tokens a day as rivals pull their free offers

Nvidia is now one of the biggest suppliers of free AI on the open market. On November 9, the free version of its largest model, Nemotron 3 Ultra, entered the top 15 on OpenRouter's public usage rankings. OpenRouter is a marketplace that lets businesses switch between hundreds of AI models through one account.

The model ranked 14th, at 3.16 trillion tokens for the day. Tokens are the small chunks of text that AI providers meter and bill. That is about 1.5% of the 208.4 trillion tokens OpenRouter's rankings show it served that day across 652 models. Nvidia charged nothing for any of it.

The rise comes as other companies withdraw free offers. For buyers, that makes Nvidia's giveaway both an opportunity and a dependency worth watching.

What Nvidia is giving away, and what it costs

Nemotron 3 Ultra launched on June 4. It is open-weight, meaning anyone can download the model and run it on their own hardware. It is large: 550 billion parameters, the internal settings a model learns, though only about 55 billion are used for each word it produces. That design makes it cheaper to run than its size suggests.

Nvidia released it under OpenMDW-1.1, a Linux Foundation license. Its model card says the model is ready for commercial use. Analysts at Digital Applied describe the terms as royalty-free, with a patent-termination clause. Nvidia also published much of the training data and its end-to-end training recipe, though some datasets remain private or gated. Few large rivals go that far.

The paid version on OpenRouter lists at $0.50 per million input tokens and $2.20 per million output tokens. Independent tester Artificial Analysis put the blended cost near $0.52 per million for typical use, according to Digital Applied. By machine.net's own rough estimate, not an Nvidia figure, one day of free traffic at the November 9 level would be worth about $1.6 million at that rate. Nvidia's actual cost to serve it is not public.

Nvidia has not publicly explained why it gives open models away. One reading, offered by Tech Insider, is that open models pull developers toward Nvidia's software and, in turn, its chips. Every company that builds on Nemotron and later moves in-house needs hardware to run it.

The free shelf is getting thinner

Nvidia's model is gaining while the rest of the free market shrinks. Independent tester TeamDay tracked OpenRouter's free text-model list from October 28 to November 6. The count fell from 16 models to 11, with none added.

Poolside's two free models expired on October 31. Thinking Machines' free Inkling models expired on November 4, according to a separate check by automation consultancy AY Automate. Both Google Gemma 4 free endpoints failed every TeamDay request with rate-limit errors.

Nemotron 3 Ultra was one of the reliable survivors. It answered all 20 of TeamDay's test requests and passed every formatting, tool-use and coding task. But it was slow in those tests, at a median of 29 tokens per second. Four of the 11 remaining free models are Nvidia's, per TeamDay, giving the company an outsized share of what is left.

Coding tools appear to drive much of the demand. OpenRouter's model page lists Kilo Code, Nous Research's Hermes Agent and Cline as top apps using the free version. Kilo Code announced on June 4 that the model would be free in its products "for a limited time." It gave no end date.

What buyers should weigh

Free does not mean private. Nvidia's free endpoints may use prompts for training, according to TeamDay's review of OpenRouter's listings. AIReiter, a developer guide, says the listing states that free-route usage is logged for security and may be used to improve Nvidia products. It advises against sending confidential or personal data. TeamDay adds that OpenRouter will not send requests to those endpoints if a customer turns off training in its privacy settings.

Free also does not mean permanent. AY Automate calls OpenRouter's free models a temporary promotion with no notice period. Both testers advise keeping a paid backup model ready. A team that wires a free model into a product could face a sudden bill or outage.

Then there is quality. Artificial Analysis scored the model 48 on its Intelligence Index at launch, ninth of 89 models, according to Digital Applied. Kilo Code called it the smartest US-made open model to date, but it trailed China's Kimi K2.6, at 54. Digital Applied also found it produced about 2.3 times as much output as the median model on that test. Since output is billed, that wordiness could offset Nvidia's own claim of up to 30% lower cost per task.

Meanwhile, Chinese open models still lead on volume. DeepSeek's V4.1 Flash alone handled 55.45 trillion tokens on November 9, or 26.6% of OpenRouter's total, per its rankings. Nvidia's free tier is a foothold, not a takeover.

The bottom line for decision makers: Nvidia is subsidizing access to a capable US-made open model while rivals step back. That is useful for prototyping and cost testing. For anything sensitive or customer-facing, the safer path is the paid route, or running the open weights on your own hardware, which the license allows.