Nemotron 3 and the open-weights math: when 'free' actually costs less

By LW Forge — maintainer of LLM Scout · Updated August 24, 2026

NVIDIA's Nemotron 3 Ultra is a 550B-parameter open mixture-of-experts model with 55B active parameters and a 1M-token context window. You can download the weights. You can also call it through an aggregator for $0.50 per million input tokens and $2.20 per million output — or for nothing at all, on a free endpoint. Three prices for one model is confusing until you notice they're answers to three different questions.

The prices, as of August 2026

Rates are OpenRouter-style aggregator pricing checked against the live listing on 2026-08-04; ours were last verified 2026-07-20 and still match. This is the first thing to internalize about open weights: there is no single price. The same checkpoint is served by a dozen hosts at different rates, different throughput, different context limits and different quantizations. A model being open doesn't fix its price; it removes the single seller who could set one. Check current rates before budgeting — our sources page lists where these come from.

The second thing: Nemotron 3 Ultra at $2.20 output is not the cheap end of open weights. Llama 4 Maverick is 3.7× cheaper on output. Nemotron 3 Super, the 120B sibling, is nearly 5× cheaper than Ultra. Ultra is priced as a reasoning-and-orchestration model, and it's the strongest U.S. open-weight entrant by most public rankings — you pay for that. If you just want the cheapest competent token, look at the cheapest LLM APIs of 2026 or the DeepSeek comparison instead.

The catch in the free tier

Aggregators list free endpoints for Nemotron 3 Ultra, Super and the Nano Omni multimodal model. These are real — real weights, real inference, $0.00 on the invoice. The price is paid elsewhere.

On OpenRouter, using free endpoints requires opting into prompt logging and training in your account's privacy settings. There are explicit toggles for "endpoints that may train on inputs" and "endpoints that may publish prompts," and free models generally sit behind them. Free tiers also come with aggressive rate limits and no availability guarantee — the endpoint that worked in your prototype can be saturated or gone the week you launch.

So the honest reading of a free tier: it's excellent for evaluation, prototypes, hobby projects, and any prompt you'd be comfortable seeing in a public dataset. It is disqualifying for customer data, anything under an NDA, health or financial records, or a product with an SLA. If your reason for looking at open weights was data governance in the first place, the free tier inverts exactly the property you wanted. Paid endpoints on the same aggregators can be filtered to providers that don't train on inputs — that's the version worth pricing.

What Ultra costs if you actually host it

Here's where the "you can run it yourself" argument meets arithmetic. Nemotron 3 Ultra at FP16 needs roughly 1.1TB of GPU memory. The standard deployment is eight GPUs with tensor parallelism across them; NVIDIA's NVFP4 checkpoint brings that down and runs on Blackwell hardware. Either way, the unit of deployment is a node, not a card.

Eight H100s at RunPod's $2.89/hr is $23.12 an hour — $16,878 a month, running or not. At the hosted rate of $2.20 per million output tokens, that same money buys about 7.7 billion output tokens. To break even you'd need to sustain roughly 2,900 output tokens per second, continuously, all month. On AWS at $6.88 per H100-hour the node runs about $40,200 a month and the breakeven moves to 18.3 billion tokens. Price your own node in the GPU cloud cost calculator; the full GPU pricing comparison has the per-provider table and the hidden costs — idle time, checkpoint storage, egress, ops — that this estimate leaves out.

The generalized version of that math: a single GPU at $2/hour running full time costs about $1,460 a month, which buys roughly 2.4 billion output tokens of Llama 4 Maverick or 660 million of Nemotron 3 Ultra. Self-hosting competes with the cheap open-weights tier and essentially never with a frontier model — and for Ultra specifically, one GPU can't even load it.

Open weights ≠ cheap API

These are two different products that happen to share a name on a leaderboard.

Open weights buy you control: run it in your own VPC or air-gapped, guarantee no third party sees a prompt, fine-tune on proprietary data, pin a version so it never changes under you, and keep serving it after the vendor loses interest. Those are governance and continuity properties. They are worth real money to regulated teams — and they are the only reasons that reliably justify the cost.

A cheap hosted API buys you the token, and nothing else. No node, no vLLM tuning, no OOM at 3am, no paying for idle GPUs overnight. At $2.20 per million output tokens with per-token billing and zero utilization risk, an aggregator beats a self-hosted node for nearly everyone below sustained industrial volume.

The practical shape most teams land on: prototype on a free endpoint with non-sensitive data, move to a paid aggregator endpoint with training disabled once real users show up, and revisit self-hosting only when compliance demands it or when your sustained token throughput is genuinely in the billions per month. Measure your actual volume in the token calculator before assuming which side of that line you're on — most teams overestimate by an order of magnitude, and the ones who don't already know it.