Nemotron 3 Ultra cost & VRAM calculator
What does Nemotron 3 Ultra cost to run for your workload, and can you run it on your own hardware? Set your workload below - we compute per-provider API cost live and tell you honestly whether local hardware can run it.
This is where owning stops making sense
Nemotron 3 Ultra is a 550-billion-parameter model. No rig an individual can buy runs it - so unlike a workstation GPU model, there's no break-even to compute. The honest answer for nearly everyone is the per-provider API cost below.
550B total, 55B active per token (MoE) - NVIDIA’s from-scratch hybrid Mamba-2 + Transformer with LatentMoE routing and 5-token Multi-Token Prediction (MTP) speculative decode. 262K default context (up to 1M with env flags); RULER@1M 94.7. Open weights under the Linux Foundation’s OpenMDW-1.1 license, downloadable on HuggingFace (BF16 + NVFP4 checkpoints). Server-class only. NVIDIA’s stated minimum is 8x H200, 16x H100, or 8x B200/GB200. BF16 weights are ~1.1TB and the NVFP4 checkpoint ~275GB, so it does not fit a single RTX 5090, RTX 6000 Ada, or Mac Studio M3/M4 Ultra. Not workstation-fittable, so cloud-only for nearly everyone despite the open weights. Benchmarks are NVIDIA self-reported: MMLU-Pro 86.8, GPQA 87.0, SWE-bench Verified 70.7. Independent eval (Artificial Analysis, Intelligence Index 47.7, #9 of 89) calls it the leading US open-weight model but ~6 points behind Kimi K2.6 on hard reasoning and coding. Treat the vendor throughput figures as claims pending community replication.
Your workload
API cost for your workload
| Provider | Rate ($/1M) | Monthly cost |
|---|---|---|
| DeepInfra | $0.50 in · $2.20 out · $0.10 cache | $0.01 cheapest |
| OpenRouter may be stale | $0.50 in · $2.20 out | $0.01 cheapest |
| BaseTen | $0.60 in · $2.40 out · $0.12 cache | $0.01 cheapest |
| Venice | $0.62 in · $3.12 out · $0.19 cache | $0.02 |
Monthly cost is an estimate from list prices and your workload - verify against the provider before committing. Cached fraction applies the cache rate to that share of input.
Can you run it locally?
Nemotron 3 Ultra has no published quantization that fits a rig one person can buy, so there's no local-hardware recommendation and no break-even to compute. The honest answer is the per-provider API cost above.
550B total, 55B active per token (MoE) - NVIDIA’s from-scratch hybrid Mamba-2 + Transformer with LatentMoE routing and 5-token Multi-Token Prediction (MTP) speculative decode. 262K default context (up to 1M with env flags); RULER@1M 94.7. Open weights under the Linux Foundation’s OpenMDW-1.1 license, downloadable on HuggingFace (BF16 + NVFP4 checkpoints). Server-class only. NVIDIA’s stated minimum is 8x H200, 16x H100, or 8x B200/GB200. BF16 weights are ~1.1TB and the NVFP4 checkpoint ~275GB, so it does not fit a single RTX 5090, RTX 6000 Ada, or Mac Studio M3/M4 Ultra. Not workstation-fittable, so cloud-only for nearly everyone despite the open weights. Benchmarks are NVIDIA self-reported: MMLU-Pro 86.8, GPQA 87.0, SWE-bench Verified 70.7. Independent eval (Artificial Analysis, Intelligence Index 47.7, #9 of 89) calls it the leading US open-weight model but ~6 points behind Kimi K2.6 on hard reasoning and coding. Treat the vendor throughput figures as claims pending community replication.
See the model card for the full architecture notes and any cloud subscription plans.