NVIDIA: Nemotron 3 Ultra.
A 550B open-weights text model (55B active) for complex agents, long-context analysis and multilingual reasoning, with up to 1M tokens of context.
2× AMD MI300X · $1.00/hr.
Most-aggressive quantisation we have a working recommendation for. Lower precision = less VRAM = cheaper hardware, at a small accuracy cost.
Cheapest hosted endpoints.
Variants in the Nemotron family.
NVIDIA: Nemotron 3.5 Lightning
Nemotron 3.5 Lightning is NVIDIA's 30B mixture-of-experts text model, 3B acti...
NVIDIA: Nemotron 3.5 Content Safety
Nemotron 3.5 Content Safety is NVIDIA's 4B moderation model that labels promp...
NVIDIA: Nemotron 3.5 Super VL
A 120B open-weights vision-language model (12B active) that reads text, image...
NVIDIA: Nemotron 3 Nano 30B A3B
NVIDIA Nemotron 3 Nano 30B A3B is a small language MoE model with highest com...
NVIDIA: Nemotron 3 Super
A 120B open-weights text model (12B active) for agent workflows, tool use and...
NVIDIA: Nemotron Nano 9B V2
NVIDIA-Nemotron-Nano-9B-v2 is a large language model (LLM) trained from scrat...
Nemotron-3-Nano-Omni-30B-A3B-Reasoning
Nemotron 3 Nano Omni is an open multimodal model built on a hybrid Mixture-of...
Cheapest hardware per quantisation.
Each row is one quantisation tier (the same weights compressed differently). Lower precision → lower VRAM → cheaper hardware, at the cost of small accuracy loss. $/hr refreshed hourly from each provider's API.
| Quantisation | Cheapest GPU config | Total VRAM | Live $/hr | tokens/sec | |
|---|---|---|---|---|---|
|
FP16
FP16 — half precision (default)
|
1536 GB | $4.00/hr | — | Compare → | |
|
FP8
FP8 — 8-bit float (Hopper / Blackwell)
|
768 GB | $2.00/hr | — | Compare → | |
|
INT4
INT4 — 4-bit integer (~4× VRAM saving)
|
384 GB | $1.00/hr | — | Compare → |
What it costs per month across providers.
Estimate your monthly bill for NVIDIA: Nemotron 3 Ultra across every host that publishes per-token pricing. Slide your token volumes; the chart + table re-rank cheapest-first.
Cheapest provider on the left.
Total monthly cost — input + output tokens combined.
Bill breakdown.
Rent the GPU instead of paying per token.
For an open-weights model like NVIDIA: Nemotron 3 Ultra, you can rent a GPU and serve inference yourself. The math: cheapest GPU rental × 730 hours/month + your electricity rate × power draw.
Assumes the GPU runs 24/7 at ~85% utilisation. If your traffic is bursty, you'll pay less for the API and probably more for the GPU (idle hours still cost rental). The breakeven analysis lives on the Self-host vs API breakeven tool.
About NVIDIA: Nemotron 3 Ultra.
An open-weights text model in NVIDIA's Nemotron 3 family, with 550B parameters and 55B active per token. It shares the family's LatentMoE design of Mamba-2, mixture-of-experts and attention layers, supports up to 1M tokens of context, and lets you turn reasoning on or off. NVIDIA recommends it for complex multi-step agents, long-context analysis, tool use, multilingual work and retrieval, and lists 8x B200-class GPUs, 8x H200 or 16x H100 as the minimum for the BF16 weights. The licence is OpenMDW-1.1.