Llama 3.3 70B

What Llama 3.3 70B is and who makes it

Llama 3.3 70B is a text model from Meta with 70 billion parameters. Meta released it on December 6, 2024 as the last update of the Llama 3 generation, a few months before Llama 4 arrived. It is a dense model, which means all of its parameters work on every token, unlike the mixture-of-experts design of Llama 4. Its weights are downloadable under the Llama 3.3 Community License.

According to Meta's model card, Llama 3.3 takes multilingual text in and produces multilingual text and code, has a context window of 128K tokens and was trained on data up to December 2023. The instruction-tuned version was aligned with supervised fine-tuning and reinforcement learning from human feedback, so it follows requests and holds a conversation well. It cannot read images.

Llama 3.3 70B is one of the four models in our chat. Llama.Online is an independent site, not run by Meta; the model runs on a cloud inference provider, and our site passes your messages to it.

Where it sits in the Llama family

The Llama 3 generation grew in steps: Llama 3 in April 2024 with 8B and 70B models, Llama 3.1 in July 2024 with 8B, 70B and 405B models and a longer 128K context, Llama 3.2 in September 2024 with small and vision models, and finally Llama 3.3 in December 2024 with this single 70B text model. Llama 4 followed in April 2025 with Scout and Maverick, which add images and use a different architecture. Many developers still use Llama 3.3 because it is predictable, well understood and supported by almost every tool. See Llama history for the whole timeline.

What Llama 3.3 70B is good at

• All-round text work. Explanations, summaries, outlines, emails and structured documents with headings and lists.
• Longer writing. Reports, essays and articles where you want a consistent tone from start to finish.
• Coding help. Explaining code, writing functions, suggesting tests and finding bugs in snippets you paste.
• Multilingual text. Eight languages are officially supported: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• A second opinion. Because it is a different model from Llama 4, asking it the same question can surface a different angle.

How to use it in our chat

1. Open the chat with the link below, or choose Llama 3.3 70B in the model menu.
2. Describe your task with context: who it is for, how long, what tone.
3. Ask for changes in follow-up messages, for example "shorter", "more formal" or "add an example".

Chat with Llama 3.3 70B

If you attach an image while Llama 3.3 is selected, the chat switches to Llama 4 Maverick, because only the Llama 4 models read images. The chat is free within a daily allowance, currently 100 messages per day.

Limits to know about

• Text only. No photos, screenshots or charts; use a Llama 4 model for those.
• Older knowledge. Its training data ends in December 2023, eight months earlier than Llama 4. It does not know about later events, and our chat has no web access.
• Fewer languages. Meta officially supports eight languages, compared with twelve for Llama 4. Other languages may work with more errors.
• Confident errors. Like any language model it can invent facts, names or citations. Check important details; our fact-checking guide helps.
• Not the newest. Newer models from other companies, and Meta's own Muse models, are generally stronger at demanding coding and multi-step agent tasks. None of the Muse models are available in our chat.

Getting Llama 3.3 70B officially

Meta publishes the weights on Hugging Face as meta-llama/Llama-3.3-70B-Instruct; accept the license on the model page to unlock the files. The model is also available in Ollama's library and in many cloud services. OpenRouter lists it at $0.10 per million input tokens and $0.32 per million output tokens (OpenRouter list price, Oct 2026), with a context window of 131,072 tokens at our provider.

Running a 70-billion-parameter model yourself needs a powerful machine with a lot of memory, typically a workstation with large graphics cards or a high-memory Apple computer. Our run Llama locally guide explains the options honestly.

The Llama 3.3 Community License allows commercial use. You must follow Meta's acceptable use policy, show "Built with Llama" in products that use it, start the names of models you train from it with "Llama", and ask Meta for a separate license if your products had more than 700 million monthly active users. Read more on our license page.

Privacy notes

In our chat, your messages go through our server and AI gateway to the inference provider so the model can answer. We do not store your message text, and your conversation history stays in your browser. Don't paste passwords or sensitive personal details into any AI chat. If you need data to stay on your own hardware, self-hosting Llama 3.3 is a common choice for companies.

How Llama 3.3 70B compares

Compared with Llama 4, it lacks images and has older knowledge, but some people prefer its steady, familiar style for long text tasks; our Llama 3.3 vs Llama 4 page weighs both. Compared with Llama 3.1 8B, it is slower but much better at nuance, longer answers and harder questions. For a broader view of open models, see Llama vs Qwen and Llama vs Mistral.

FAQ

Is Llama 3.3 70B better than Llama 4?

Not overall. Meta says Llama 4 Maverick offers higher quality at a lower price than Llama 3.3 70B. Still, Llama 3.3 remains a solid text model and is sometimes preferred for its predictable behavior.

Can Llama 3.3 read images?

No, it is text only. In our chat, attaching an image switches you to Llama 4 Maverick.

What is the knowledge cut-off of Llama 3.3?

Meta's model card lists December 2023.

Can I use Llama 3.3 commercially?

Yes, under the Llama 3.3 Community License and its acceptable use policy, with the "Built with Llama" requirement and the 700 million user rule. This is general information, not legal advice.

Is there a Llama 3.3 8B?

Meta's Llama 3.3 model card describes only the 70B model. For a small, fast Llama, use Llama 3.1 8B.