# Llama 3.3 Nemotron Super 49B v1 ![Organization](https://img.shields.io/badge/Org-NVIDIA-76b900) ![Parameters](https://img.shields.io/badge/Params-49B-1f6feb) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Llama%203.3-d29922) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Apr%202025-555) > **Quick answer:** Llama 3.3 Nemotron Super 49B v1 is NVIDIA's April 2025 distilled and aligned 49B model, scoring 88.3% Arena Hard, 73.7% BFCL v2, and 9.17/10 MT-Bench. Built from Llama 3.3 70B via NVIDIA's distillation, it achieves higher MT-Bench than the 70B base model while using fewer parameters. ## At a Glance **Where Nemotron Super 49B leads** - 9.17 MT-Bench — highest open-weight MT-Bench score among Nemotron models - 88.3% Arena Hard — excellent instruction following - 73.7% BFCL v2 — strong tool use and function calling - 49B: smaller than 70B, more inference-efficient - NVIDIA's advanced alignment (RLHF + distillation) **Where it lags** - Llama 3.3 License — commercial restrictions apply - Text-only; no vision - Smaller than Nemotron Ultra 253B for maximum performance tasks - August 2024 knowledge cutoff **Best for:** High-quality instruction following and tool use at 49B scale; agentic pipelines requiring strong function calling; NVIDIA GPU deployments. ## What Llama 3.3 Nemotron Super 49B Is Llama 3.3 Nemotron Super 49B v1 (released April 14, 2025) is NVIDIA's "Super" tier in the Nemotron family — built on Llama 3.3 70B and reduced to 49B via knowledge distillation, then further aligned with NVIDIA's RLHF pipeline. The 9.17 MT-Bench score is exceptional: it exceeds the base Llama 3.3 70B while using fewer parameters, demonstrating that NVIDIA's alignment improvements can compensate for the parameter reduction. The 73.7% BFCL v2 score is also notable for agentic use cases. NVIDIA offers three Nemotron tiers: Nano (8B), Super (49B), and Ultra (253B). Super targets the sweet spot of high alignment quality with manageable inference cost. ## Specifications | Field | Value | |---|---| | Organization | NVIDIA | | Parameters | 49B (dense) | | Context window | 128,000 tokens | | License | Llama 3.3 License | | HuggingFace | [nvidia/Llama-3.3-Nemotron-Super-49B-v1](https://huggingface.co/nvidia/Llama-3.3-Nemotron-Super-49B-v1) | | Base model | Meta Llama 3.3 70B (distilled to 49B) | | Release date | April 14, 2025 | | Knowledge cutoff | August 2024 | | Modality | Text only | ## Pricing Open weights under Llama 3.3 License. Available via NVIDIA API Catalog (build.nvidia.com). ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [Arena Hard](/benchmarks/arena-hard-v2) | 88.3% | Benchgen evaluation | 2025-07 | | [BFCL v2](/benchmarks/bfcl-v3) | 73.7% | Benchgen evaluation | 2025-07 | | [MT-Bench](/benchmarks/mt-bench) | 9.17 / 10 | Benchgen evaluation | 2025-07 | ## Nemotron Super 49B vs Alternatives | Model | MT-Bench | Arena Hard | BFCL v2 | Params | |---|---|---|---|---| | **Nemotron Super 49B** | 9.17 | 88.3% | 73.7% | 49B | | [Llama 3.1 Nemotron 70B](/models/nvidia/llama-3-1-nemotron-70b-instruct) | 9.0 | — | — | 70B | | [Llama 3.3 70B Instruct](/models/meta/llama-3-3-70b-instruct) | — | — | — | 70B | | [Nemotron Ultra 253B](/models/nvidia/llama-3-1-nemotron-ultra-253b-v1) | — | — | 74.1% | 253B | Nemotron Super 49B vs Nemotron 70B: higher MT-Bench (9.17 vs 9.0) at fewer parameters (49B vs 70B). vs Nemotron Ultra 253B: similar BFCL but much lower inference cost. ## Related Models & Benchmarks - Compare: [Llama 3.1 Nemotron 70B](/models/nvidia/llama-3-1-nemotron-70b-instruct) · [Nemotron Ultra 253B](/models/nvidia/llama-3-1-nemotron-ultra-253b-v1) · [Llama 3.3 70B Instruct](/models/meta/llama-3-3-70b-instruct) - Benchmarks: [MT-Bench](/benchmarks/mt-bench) · [Arena Hard](/benchmarks/arena-hard-v2) · [BFCL v3](/benchmarks/bfcl-v3) ## Frequently Asked Questions

What is Llama 3.3 Nemotron Super 49B?

NVIDIA's April 2025 distilled 49B model scoring 88.3% Arena Hard, 73.7% BFCL v2, and 9.17 MT-Bench. Built from Llama 3.3 70B via distillation and NVIDIA's alignment pipeline.

Is Nemotron Super 49B open source?

The weights are available on Hugging Face under the Llama 3.3 License. Commercial use permitted with some restrictions.

--- *Specs from NVIDIA's official Nemotron Super 49B release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.*

Llama-3.3 Nemotron Super 49B v1 — Evaluations

Evaluations