Model Details
Llama 3.1 Nemotron Ultra 253B
Quick answer: Llama 3.1 Nemotron Ultra 253B is NVIDIA's April 2025 post-trained model built on Meta's Llama 3.1 405B base, scoring 74.1% BFCL v2 and 66.3% LiveCodeBench. It is NVIDIA's largest Nemotron model — above Nemotron Super 49B (73.7% BFCL v2) in function calling.
At a Glance
Where Nemotron Ultra 253B leads
- 74.1% BFCL v2 — strong function calling, above Nemotron Super 49B (73.7%)
- 66.3% LiveCodeBench — competitive coding performance at this size
- Built on Llama 3.1 405B base — full Meta ecosystem compatibility
- 128K context window (via Llama 3.1)
Where it lags
- 253B parameters: requires significant GPU memory (~500GB+ for inference)
- Llama 3.1 license restrictions (not fully permissive for all commercial uses)
- Less benchmark data than OpenAI/Google frontier models
- April 2025 release — pre-dates some newer SOTA
Best for: Enterprise deployments requiring large-scale reasoning with NVIDIA optimisation; tool-use and function-calling agent pipelines; teams using NVIDIA infrastructure.
What Llama 3.1 Nemotron Ultra 253B Is
Llama 3.1 Nemotron Ultra 253B is NVIDIA's April 2025 post-training of Meta's Llama 3.1 405B. NVIDIA's Nemotron series involves RLHF and preference optimisation on top of Meta's base models, producing variants that excel on tool-use, instruction-following, and agentic tasks.
"Ultra" indicates the top tier in NVIDIA's Nemotron lineup for this generation, with "Super" (49B) being the mid-tier variant. At 253B parameters on Llama 3.1 base, it inherits the 128K context window while benefiting from NVIDIA's alignment training.
Specifications
| Field | Value |
|---|---|
| Organization | NVIDIA |
| License | Llama 3.1 License |
| HuggingFace | nvidia/Llama-3.1-Nemotron-Ultra-253B-v1 |
| Release date | April 14, 2025 |
| Base model | Llama 3.1 405B |
| Parameters | 253B |
| Modality | Text only |
| Context window | 128K tokens |
Pricing
Open weights (Llama 3.1 License) — self-host via NVIDIA NIM or HuggingFace. Available via NVIDIA API Catalog.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| BFCL v2 | 74.1% | Benchgen evaluation | 2025-04 |
| LiveCodeBench | 66.3% | Benchgen evaluation | 2025-04 |
Nemotron Ultra 253B vs Alternatives
| Model | BFCL v2 | LiveCodeBench | Params | License |
|---|---|---|---|---|
| Nemotron Ultra 253B | 74.1% | 66.3% | 253B | Llama 3.1 |
| Nemotron Super 49B | 73.7% | — | 49B | Llama 3.3 |
| Llama 3.1 405B Instruct | — | — | 405B | Llama 3.1 |
| QwQ-32B | 66.4% | — | 32B | Apache 2.0 |
Nemotron Ultra 253B vs Super 49B: marginal BFCL gain (74.1% vs 73.7%) at ~5x the parameter cost. Super 49B is more efficient for most deployments.
Related Models & Benchmarks
- Compare: Nemotron Super 49B · Llama 3.1 405B Instruct · QwQ-32B
- Benchmarks: BFCL v3 · LiveCodeBench
Frequently Asked Questions
What is Llama 3.1 Nemotron Ultra 253B?
NVIDIA's April 2025 post-training of Llama 3.1 405B, scoring 74.1% BFCL v2 and 66.3% LiveCodeBench. NVIDIA's largest Nemotron model optimised for function-calling and agentic tasks.Is Nemotron Ultra 253B better than Nemotron Super 49B?
Marginally on BFCL (74.1% vs 73.7%) at ~5x parameter cost. For most deployments, Nemotron Super 49B offers better efficiency. Choose Ultra 253B only if you need maximum BFCL-class function calling and have the hardware.Specs from NVIDIA's Llama 3.1 Nemotron Ultra 253B release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.