# Llama 3.3 70B Instruct ![Organization](https://img.shields.io/badge/Org-Meta-0064e0) ![Parameters](https://img.shields.io/badge/Params-70B-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-Dec%202024-555) > **Quick answer:** Llama 3.3 70B Instruct is Meta's December 2024 improved 70B model, scoring 96.0% on AI2 Reasoning Challenge, 86.0% on MMLU, 77.0% on MATH, 88.4% on HumanEval, and 68.9% on MMLU-Pro. It significantly improves on Llama 3.1 70B — notably +27pp MATH (77% vs 50%) — while maintaining the same 128K context window. ## At a Glance **Where Llama 3.3 70B leads** - 77.0% MATH — strong competition mathematics, +27pp vs Llama 3.1 70B - 88.4% HumanEval — excellent code completion - 68.9% MMLU-Pro — improved academic knowledge - 96.0% AI2 Reasoning Challenge - 128K context window - Open weights with strong ecosystem support **Where it lags** - Llama 3.3 License (not Apache 2.0) — commercial restrictions - Text-only: no vision capability - Superseded by Llama 4 Maverick for multimodal and 1M context needs - August 2024 knowledge cutoff **Best for:** Open-weight 70B deployments requiring strong math/coding; teams on Llama 3.1 70B seeking improvements without switching to Llama 4; complex STEM tasks. ## What Llama 3.3 70B Instruct Is Llama 3.3 70B Instruct (released December 6, 2024) is Meta's iterative improvement to the Llama 3.1 70B architecture, trained with enhanced data and improved instruction tuning. The most notable improvement is MATH: 77.0% vs approximately 50% for Llama 3.1 70B — a 27-percentage-point gain on competition mathematics. The model achieves 88.4% HumanEval (vs ~72% for 3.1 70B) and 68.9% MMLU-Pro, placing it solidly in the frontier open-weight tier at the 70B scale. It maintains the same 128K context window and architecture, making it a drop-in upgrade for most Llama 3.1 70B deployments. For teams choosing between Llama 3.1 70B and 3.3 70B, Llama 3.3 is the clear choice: substantially better math, coding, and knowledge with minimal migration overhead. ## Specifications | Field | Value | |---|---| | Organization | Meta | | Parameters | 70B (dense) | | Context window | 128,000 tokens | | License | Llama 3.3 License | | HuggingFace | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) | | Release date | December 6, 2024 | | Knowledge cutoff | August 2024 | | Modality | Text only | ## Pricing Open weights under Llama 3.3 License on Hugging Face. Available via hosted providers (Together AI, Groq, etc.) at market rates. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [AI2 Reasoning Challenge](/benchmarks/ai2-reasoning-challenge) | 96.0% | Benchgen evaluation | 2025-07 | | [MMLU](/benchmarks/mmlu) | 86.0% | Benchgen evaluation | 2025-07 | | [MATH](/benchmarks/math) | 77.0% | Benchgen evaluation | 2025-07 | | [HumanEval](/benchmarks/humaneval) | 88.4% | Benchgen evaluation | 2025-07 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 68.9% | Benchgen evaluation | 2025-07 | | [BigCodeBench](/benchmarks/bigcodebench) | 28.4% | Benchgen evaluation | 2025-07 | ## Llama 3.3 70B vs Alternatives | Model | MATH | HumanEval | MMLU-Pro | License | |---|---|---|---|---| | **Llama 3.3 70B Instruct** | 77.0% | 88.4% | 68.9% | Llama 3.3 | | [Llama 3.1 70B Instruct](/models/meta/llama-3-1-70b-instruct) | ~50% | ~72% | — | Llama 3.1 | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | — | — | 80.5% | Llama 4 | | [DeepSeek-V3](/models/deepseek/deepseek-v3) | 90.2% | — | — | MIT | Llama 3.3 70B vs 3.1 70B: +27pp MATH, +16pp HumanEval — a clear upgrade with minimal migration overhead. vs Llama 4 Maverick: 4 Maverick wins on MMLU-Pro (80.5% vs 68.9%) and context (1M vs 128K) but requires a MoE-aware inference stack. ## Related Models & Benchmarks - Compare: [Llama 3.1 70B Instruct](/models/meta/llama-3-1-70b-instruct) · [Llama 4 Maverick](/models/meta/llama-4-maverick) · [DeepSeek-V3](/models/deepseek/deepseek-v3) - Benchmarks: [MATH](/benchmarks/math) · [HumanEval](/benchmarks/humaneval) · [MMLU-Pro](/benchmarks/mmlu-pro) ## Frequently Asked Questions

What is Llama 3.3 70B Instruct?

Llama 3.3 70B Instruct is Meta's December 2024 improved 70B model, scoring 96.0% AI2 RC, 86.0% MMLU, 77.0% MATH, and 88.4% HumanEval with 128K context. It significantly outperforms Llama 3.1 70B on mathematics and coding.

Should I use Llama 3.3 70B or Llama 3.1 70B?

Llama 3.3 70B for new deployments: +27pp MATH (77.0% vs ~50%), +16pp HumanEval — a significant upgrade with minimal migration overhead as it uses the same architecture.

Should I use Llama 3.3 70B or Llama 4 Maverick?

Llama 4 Maverick achieves better MMLU-Pro (80.5% vs 68.9%), 1M context, and multimodal capability. Llama 3.3 70B is preferred when a dense 70B architecture is specifically required.

--- *Specs from Meta's official Llama 3.3 announcement (December 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Llama 3.3 70B Instruct — Evaluations

Evaluations