Nemotron 3 Super
NVIDIA's open-weight model for coordinating agents, planning work, and combining their results, served on Boundless.
About the model
Nemotron 3 Super at a glance
Nemotron 3 Super is NVIDIA's 120B hybrid Mamba-Transformer mixture-of-experts model, with 12B active per pass. It is built to coordinate multi-agent systems by routing work, planning and combining results, and its weights, training data and recipes are open source.
From our model docs- Developer
- NVIDIA
- Context window
- 262K tokens
- Input
- Text
- Capabilities
- Reasoning, tool use
Tailored inference
Want Nemotron 3 Super tuned to your workload?
Our engineers benchmark it against your current model, tune how it's served for your traffic, and run it to your cost, latency and quality targets.
Talk to an engineerRun it
Switch with a base URL and a key
- 01
- 02
Point your SDK at Boundless
Change the base URL and key. For Anthropic SDKs, drop the /v1 and set max_tokens.
curl https://api.inference.boundless.network/v1/chat/completions \
-H "Authorization: Bearer $BOUNDLESS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nemotron3-super",
"messages": [{
"content": "Say hello",
"role": "user"
}]
}'
FAQ
Questions before you commit
How much does Nemotron 3 Super cost on Boundless?
$0.12 per 1M input tokens, $0.06 per 1M cached input tokens and $0.52 per 1M output tokens, billed from prepaid credits. Every response includes exact token counts, so any charge can be checked against these prices.
What is Nemotron 3 Super's context window on Boundless?
262K tokens, shared by the prompt and the reply. max_tokens can go up to that limit.
Does Nemotron 3 Super support tool calling?
Yes. Nemotron 3 Super supports tool use and reasoning.
Can I control how much Nemotron 3 Super reasons?
Yes. On chat completions, set reasoning_effort to none, low, medium or high, or leave it unset for the model's default. Size max_tokens for both the thinking and the reply.
What are the rate limits?
Each team has caps on tokens per minute, requests per minute and concurrent requests, shown on the usage page. Cached input doesn't count against them, and support can raise them for your team.
Does Boundless store my prompts or outputs?
No. Boundless never stores prompts or outputs and never trains on them. It keeps only what billing needs: token counts, model, timestamp and cost.
Can Boundless help me decide if Nemotron 3 Super fits my workload?
Yes. We test it against your current model on your own prompts, then tune and run it to your targets.
More models
