nvidia
NVIDIA: Nemotron 3 Ultra
nvidia/nemotron-3-ultra-550b-a55b
NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...
- Type
- text
- Provider
- nvidia
- Model ID
- nvidia/nemotron-3-ultra-550b-a55b
- Capabilities
- tools, reasoning
- Context window
- 262K tokens
- Self-hostable
- Yes — runs on your own GPU
- Input price
- $0.55 / 1M tokens
- Output price
- $2.42 / 1M tokens
Cloud price is billed from prepaid credits when a request fails over to the cloud. Open-weight models run free on GPUs you own — the gateway routes to your nodes first.
What it costs to run
cloud · 1M in + 1M out
$2.97
$0.55 for a million input tokens plus $2.42 for a million output tokens, billed from credits only when a request fails over to the cloud.
your hardware
$0 / token
Served from a workstation or on-prem server you own, there is no per-token fee — the marginal cost is power.
Run it on your own hardware
Estimated GPU memory for NVIDIA: Nemotron 3 Ultra (550B parameters), for the weights alone — add room for the context cache.
- Q4_K_M
- ~313.2 GBmulti-GPU or large unified memory
- Q8_0
- ~543.7 GBmulti-GPU or large unified memory
- FP16 / BF16
- ~1025.2 GBmulti-GPU or large unified memory
Install the agent on the machine and it dials out to the gateway over a single WebSocket — no port forwarding or inbound firewall rules. Requests for this model go to your node first and fail over to the cloud only when it can't serve.
/// initialize
Call NVIDIA: Nemotron 3 Ultra through one OpenAI-compatible endpoint.
Serve it from hardware you own, with cloud failover when your nodes are busy or offline.
no credit card · 2 nodes free · openai-compatible
Frequently asked questions
- What is the context window of NVIDIA: Nemotron 3 Ultra?
- NVIDIA: Nemotron 3 Ultra accepts up to 262,144 tokens (262K) of context per request.
- How much does NVIDIA: Nemotron 3 Ultra cost per million tokens?
- Through Wide Area Intelligence, NVIDIA: Nemotron 3 Ultra costs $0.55 per 1M input tokens and $2.42 per 1M output tokens when served from the cloud, billed from prepaid credits. A workload of 1M input plus 1M output tokens costs $2.97.
- Can I run NVIDIA: Nemotron 3 Ultra on my own hardware?
- Yes — NVIDIA: Nemotron 3 Ultra is an open-weight model you can run on a workstation or on-prem server you own. At Q4_K_M it needs roughly 313.2 GB of GPU memory, plus room for the context cache — more than a single 80 GB card, so plan for multiple GPUs or a large unified-memory machine. Wide Area Intelligence routes requests to your own nodes first and fails over to the cloud only when they can't serve.