Nemotron Ultra 253B
NVIDIA's largest open-weight dense model, a reasoning-focused flagship served through NVIDIA NIM and tuned to run efficiently on its own accelerators.
At a glance
Released
Mar 5, 2026
Provider
NVIDIA
Context window
256,000 tokens
Max output
32,000 tokens
Modalities
Text → text
Knowledge cutoff
December 2025
API model ID
nemotron-ultra-253bLicense
Open weights (NVIDIA Open Model License)
Pricing
Standard
Input$3.50 / 1M
Output$14 / 1M
* Rates are from a third-party aggregator — verify before relying on them. A typical 1,500-token prompt with a 650-token answer costs ≈ $0.014 on the standard tier. Run your own prompt through the calculator →
Best use cases
Advanced reasoning
Tackles math, science, and multi-step planning problems with long deliberation.
On-prem deployment
Runs open weights inside regulated or air-gapped NVIDIA environments.
Synthetic data
Generates high-quality training data and reasoning traces for smaller models.
Strengths & limitations
Strengths
- Strong open-weight reasoning
- Optimized for NVIDIA hardware
- Long context
Limitations
- Heavy and costly to serve
- Text only
How to test it
- No-code: open build.nvidia.com (NVIDIA NIM) and run a prompt in the playground.
- API: create a key in build.nvidia.com (NVIDIA NIM) and call the API (example below).
curl https://integrate.api.nvidia.com/v1/chat/completions \
# OpenAI-compatible endpoint
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nemotron-ultra-253b",
"messages": [{"role": "user", "content": "Explain tokenization in one paragraph."}]
}'