Model Details
Llama 3.1 Nemotron 70B Instruct
Quick answer: Llama 3.1 Nemotron 70B Instruct is NVIDIA's October 2024 instruction-aligned fine-tune of Llama 3.1 70B, scoring 91.4% on GSM8K, 85.6% on MMLU, and 9.0/10 on MT-Bench. The 9.0 MT-Bench score was notable — among the highest for an open-weight model at launch — demonstrating NVIDIA's RLHF alignment expertise.
At a Glance
Where Nemotron 70B leads
- 9.0/10 MT-Bench — among the highest open-weight instruction following scores
- 91.4% GSM8K — strong math reasoning (+7pp vs base Llama 3.1 70B)
- Excellent alignment quality: balanced, helpful responses
- NVIDIA RLHF and alignment research applied to Llama 3.1 base
- 128K context window
Where it lags
- Llama 3.1 Community License — commercial restrictions
- Superseded by Llama 3.3 70B (better benchmarks overall) and Nemotron Super/Ultra
- December 2023 knowledge cutoff — earlier than Llama 3.3 70B
- Text-only
Best for: High-quality instruction following at 70B scale; teams prioritising alignment and response quality; NVIDIA GPU deployments.
What Llama 3.1 Nemotron 70B Instruct Is
Llama 3.1 Nemotron 70B Instruct is NVIDIA's fine-tuned variant of Meta's Llama 3.1 70B, applying NVIDIA's RLHF (Reinforcement Learning from Human Feedback) and alignment techniques. Released October 15, 2024, it was designed to demonstrate how post-training alignment can improve instruction following over the base model.
The 9.0 MT-Bench score represented a significant achievement — surpassing GPT-4-level MT-Bench scores while remaining open-weight. The RLHF process improves response quality, instruction adherence, and helpfulness compared to the base Llama 3.1 70B (8.70 MT-Bench).
NVIDIA has since released Nemotron Super 49B (April 2025) with even higher MT-Bench (9.17) at fewer parameters, and Nemotron Ultra 253B for maximum performance. The Nemotron 70B remains relevant for teams that have deployed it in production.
Specifications
| Field | Value |
|---|---|
| Organization | NVIDIA |
| Parameters | 70B (dense) |
| Context window | 128,000 tokens |
| License | Llama 3.1 Community License |
| HuggingFace | nvidia/Llama-3.1-Nemotron-70B-Instruct-HF |
| Base model | Meta Llama 3.1 70B |
| Release date | October 15, 2024 |
| Knowledge cutoff | December 2023 |
| Modality | Text only |
Pricing
Open weights under Llama 3.1 Community License. Available via NVIDIA API Catalog (build.nvidia.com).
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| GSM8K | 91.4% | Benchgen evaluation | 2025-07 |
| MMLU | 85.6% | Benchgen evaluation | 2025-07 |
| HellaSwag | 85.6% | Benchgen evaluation | 2025-07 |
| MT-Bench | 9.0 / 10 | Benchgen evaluation | 2025-07 |
Nemotron 70B vs Alternatives
| Model | MT-Bench | MMLU | GSM8K | License |
|---|---|---|---|---|
| Llama 3.1 Nemotron 70B | 9.0 | 85.6% | 91.4% | Llama 3.1 |
| Llama 3.1 70B Instruct | 8.70 | 86.0% | — | Llama 3.1 |
| Llama 3.3 70B Instruct | — | 86.0% | — | Llama 3.3 |
| Nemotron Super 49B | 9.17 | — | — | Llama 3.3 |
Nemotron 70B vs Llama 3.1 70B: +0.3 MT-Bench (9.0 vs 8.70) from NVIDIA's alignment. Nemotron Super 49B achieves 9.17 MT-Bench at smaller parameter count. For new deployments, Nemotron Super 49B or Llama 3.3 70B are preferred.
Related Models & Benchmarks
- Compare: Llama 3.1 70B Instruct · Llama 3.3 70B Instruct · Nemotron Super 49B
- Benchmarks: MT-Bench · MMLU · GSM8K
Frequently Asked Questions
What is Llama 3.1 Nemotron 70B?
NVIDIA's October 2024 RLHF fine-tune of Llama 3.1 70B, scoring 91.4% GSM8K, 85.6% MMLU, and 9.0 MT-Bench with 128K context. Notable for high instruction following quality.What makes Nemotron 70B different from Llama 3.1 70B?
NVIDIA applied additional RLHF alignment training, improving MT-Bench from 8.70 to 9.0. The base model architecture and weights are identical — only the post-training alignment differs.Specs from NVIDIA's official Nemotron 70B release (October 2024) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.