# Llama 3.1 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%203.1-d29922) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Jul%202024-555) > **Quick answer:** Llama 3.1 70B Instruct is Meta's July 2024 70B instruction-tuned open-weight model. It scores 94.8% on AI2 Reasoning Challenge, 86.0% on MMLU, 8.70/10 on MT-Bench, and 25.4% on BigCodeBench. With a 128K-token context window under the Llama 3.1 License, it was the leading open-weight model at its scale at launch. ## At a Glance **Where Llama 3.1 70B leads** - 94.8% AI2 Reasoning Challenge — strong scientific reasoning - 86.0% MMLU — GPT-4-class academic knowledge at open-weight pricing - 8.70/10 MT-Bench — high-quality instruction following - 128K context window - Open weights — self-hostable under Llama 3.1 License - Foundation for many fine-tuned derivatives (Hermes, Nemotron, etc.) **Where it lags** - Superseded by Llama 4 Maverick and Scout for general use - 25.4% BigCodeBench — limited complex coding performance - Text-only: no vision capability - March 2024 knowledge cutoff - Dense 70B requires significant GPU VRAM for full-precision inference **Best for:** Self-hosted 70B-class deployment; base for fine-tuning; existing integrations built on Llama 3.1 70B; teams requiring Llama 3.1's specific instruction style. ## What Llama 3.1 70B Instruct Is Llama 3.1 70B Instruct was Meta's July 2024 flagship open-weight model, released as part of the Llama 3.1 family alongside 8B and 405B variants. It was the first Llama model with a 128K context window, and at launch, its MMLU score (86.0%) was competitive with GPT-4 class proprietary models. The model became the standard reference point for open-weight 70B performance in 2024, serving as the base for dozens of fine-tuned variants including Hermes 3 70B, Llama 3.1 Nemotron 70B, and others. Its strong instruction following (8.70 MT-Bench) and academic knowledge make it a reliable general-purpose open-weight model. For new deployments in 2025+, Llama 4 Maverick provides better performance with 1M context and multimodal capability. Llama 3.1 70B Instruct remains relevant for teams with validated production workflows, specific fine-tuning needs, or hardware constraints that suit a dense 70B architecture. ## Specifications | Field | Value | |---|---| | Organization | Meta | | Parameters | 70B (dense) | | Context window | 128,000 tokens | | License | Llama 3.1 License | | HuggingFace | [meta-llama/Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct) | | Release date | July 23, 2024 | | Knowledge cutoff | March 2024 | | Modality | Text only | ## Pricing Llama 3.1 70B Instruct is available as open weights — free to self-host under the Llama 3.1 License. Available via numerous hosted providers (Together AI, Groq, Fireworks, etc.) at market rates. ## Context Window Llama 3.1 70B Instruct has a **128,000-token context window** — roughly **90 pages** of text in a single request. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [AI2 Reasoning Challenge](/benchmarks/ai2-reasoning-challenge) | 94.8% | Benchgen evaluation | 2024-07 | | [MMLU](/benchmarks/mmlu) | 86.0% | Benchgen evaluation | 2024-07 | | [MT-Bench](/benchmarks/mt-bench) | 8.70 / 10 | Benchgen evaluation | 2024-07 | | [BigCodeBench](/benchmarks/bigcodebench) | 25.4% | Benchgen evaluation | 2024-07 | ## Llama 3.1 70B vs Alternatives | Model | Context | MMLU | MT-Bench | Weights | |---|---|---|---|---| | **Llama 3.1 70B Instruct** | 128K | 86.0% | 8.70 | Open | | [Llama 3.1 405B Instruct](/models/meta/llama-3-1-405b-instruct) | 128K | 88.6% | — | Open | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | 1M | — | — | Open | | [Hermes 3 70B](/models/nousresearch/hermes-3-70b) | 128K | — | — | Open | | [Phi-4](/models/microsoft/phi-4) | 16K | — | — | Open | Llama 3.1 70B vs 405B: 405B achieves 88.6% MMLU (vs 86.0%) at 6× more parameters and infrastructure cost. For most tasks, 70B provides excellent performance at significantly lower self-hosting cost. ## Related Models & Benchmarks - Compare: [Llama 3.1 405B Instruct](/models/meta/llama-3-1-405b-instruct) · [Llama 4 Maverick](/models/meta/llama-4-maverick) · [Hermes 3 70B](/models/nousresearch/hermes-3-70b) - Benchmarks: [AI2 Reasoning Challenge](/benchmarks/ai2-reasoning-challenge) · [MMLU](/benchmarks/mmlu) · [MT-Bench](/benchmarks/mt-bench) ## Frequently Asked Questions

What is Llama 3.1 70B Instruct?

Llama 3.1 70B Instruct is Meta's July 2024 open-weight model, scoring 94.8% AI2 RC, 86.0% MMLU, and 8.70 MT-Bench with a 128K context window under the Llama 3.1 License.

Is Llama 3.1 70B open source?

The weights are open under the Llama 3.1 License on Hugging Face. The license permits commercial use with some restrictions — review terms for your specific use case.

What is Llama 3.1 70B's context window?

Llama 3.1 70B Instruct supports a 128,000-token context window.

Should I use Llama 3.1 70B or Llama 4 Maverick?

For new deployments, Llama 4 Maverick is preferred: better benchmark scores, 1M context, and multimodal capability. Llama 3.1 70B is appropriate for existing integrations or hardware setups optimised for dense 70B models.

--- *Specs from Meta's official Llama 3.1 announcement (July 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Llama 3.1 70B Instruct — Evaluations

Evaluations