Robots Atlas>ROBOTS ATLAS
Llama 3.3 70B
AI Modelsโ€บLlama

Llama 3.3 70B

3.3 70Bย ยทย Family: Llama
Meta's open (Llama 3.3 Community License) language model: 70B parameters (dense, text-only), multilingual (8 languages), 128K context; quality comparable to Llama 3.1 405B at a much smaller size.
โœ“ Activeโœ“ Public accessโš– Open weightsLLMTool-using model๐Ÿ“ Llama
Context window
128K
tokens
Parameters
70B (dense)
parameters
Max output
2,048
tokens
Release date
6 December 2024
๐ŸขMeta AIProducer
Access:DownloadAPIHostedDeployment:๐Ÿ’ป Localโ˜ Cloud

Overview

Llama 3.3 70B is Meta's open, multilingual language model, introduced on December 6, 2024. It is a dense 70B-parameter, instruction-tuned model, text-only (text input and output, code generation).

The model uses an optimized transformer architecture with Grouped-Query Attention (GQA) for better inference scalability, supports a 128K-token context and 8 languages (including English, German, French, Spanish, Italian, Portuguese, Hindi and Thai). It features strong reasoning, coding, function calling and multi-step tasks.

A key advantage of Llama 3.3 70B is efficiency: it reaches quality close to the much larger Llama 3.1 405B (e.g. MMLU CoT ~86) at a fraction of the inference cost. Weights are released under the Llama 3.3 Community License (open weights with restrictions), allowing self-hosting; the model is available via Hugging Face and provider APIs (e.g. OpenRouter). Knowledge cutoff: December 2023.

Classification
LLMTool-using model
Family: Llama
Applications
CodingChatbotsQ&A / Question answeringKnowledge workResearch assistanceCreative writing
Access & deployment
DownloadAPIHosted
LocalCloud
Weights: Open weights
Key parameters
๐Ÿ“ Context: 128K
๐Ÿงฉ Parameters: 70B (dense)
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text
Platforms
Hugging Face HubOpenRouter

Technical specification

Context window
128K
tokens
Parameters
70B (dense)
parameters
Max output tokens
2,048
tokens per response
Knowledge cutoff
1 Dec 2023
Knowledge boundary
License
Llama 3.3 Community License
Hardware requirements
70B dense - self-hosting requires a GPU server (e.g. ~140 GB VRAM in BF16, or INT4/INT8 quantization on single/dual GPUs). Supported: Transformers, vLLM, llama.cpp, Ollama.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
text
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Function Calling
Category: planning
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Planning
Forming and executing action plans for complex tasks.
Category: planning
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language
Application domains
CodingChatbotsQ&A / Question answeringKnowledge workResearch assistanceCreative writing

Benchmark results

1 benchmark
MMLU (CoT, 0-shot)
macro_avg / accuracy ยท Llama 3.3 70B Instruct scored 86.0 on MMLU (CoT), on par with Llama 3.1 405B (88.6) per Meta's model card.
86.0%
๐Ÿ“… 6 Dec 2024
Source: Hugging Face model card (Meta).

Pricing

Technical architecture

Core Architecture
TRTransformerGQGQA
Model Form
LLLLM
Training Techniques
ITInstruction Tuning
Related Technologies
FCFunction Calling

Deployment and security

๐Ÿ”’ Security / Enterprise
โœ“ Verified enterprise information

Open weights (Llama 3.3 Community License) - self-hosting / on-premises possible with full data control. The license imposes restrictions (e.g. entities with >700M MAU). Supports function calling and tool integration.

Updated: 31 Jul 2026โ†— Security documentation

Sources and related pages