Llama 3.3 70B: Capabilities, Hardware, and When to Use It

Llama 3.3 70B is the 70-billion-parameter instruction-tuned model Meta released on December 6, 2024. Meta's model card describes a dense model with a 128K-token context window that takes text and returns text and code, in eight officially supported languages. Meta publishes the weights under its own community license.

It is the oldest model we offer, and it lacks two features the newer ones have: it has no reasoning mode and no image input. What it does have is strong instruction following in Meta's results, support in common open-source runtimes such as vLLM and Ollama, and a place on Maple's Free plan.

This guide covers what the model is, what it takes to run yourself, how to use it in Maple, and when another model is the better pick.

What is Llama 3.3 70B?

It is a conventional dense transformer: no experts are switched in and out, so each of its 80 layers runs in full for every token. That makes it heavier per token than mixture-of-experts models, and all 70.6 billion parameters have to be loaded. Its Hugging Face listing names Llama 3.1 70B as the base model; Llama 3.3 is a new round of post-training on that base, with supervised fine-tuning and reinforcement learning from human feedback.

It has no reasoning mode. Reasoning models can spend extra tokens working through a problem before they give the final answer, and they are built for multi-step math and code. Llama 3.3 70B has no such mode, in Meta's release or in Maple, so its reply is the answer itself.

Meta lists eight supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. The model saw more languages in training, but Meta's card discourages using others without fine-tuning and your own safeguards. Its pretraining data stops in December 2023.

It calls tools. Meta's prompt format recommends zero-shot function calling: you describe your functions in JSON, and the model replies with a list of calls such as [get_weather(city='Seattle')], several at once if needed.⁵

Is Llama 3.3 70B open source?

It is open-weight. You can download, run, modify and build on the weights commercially, but under Meta's own Llama 3.3 Community License, not a standard open-source license. Meta does not publish the training data.

The license adds conditions that a permissive license like Apache 2.0 does not:

• Acceptable use. Your use must follow Meta's acceptable use policy, which rules out, for example, the unlicensed practice of law or medicine and generating malware.
• Attribution. If you distribute or make available the model, a derivative, or a product or service that contains it, you include a copy of the license and show a "Built with Llama" notice. A model you build or improve with Llama or its outputs, and then distribute or make available, must have a name that starts with "Llama".
• Very large services. A company whose products and services, with its affiliates, had more than 700 million monthly active users in the month before the December 2024 release must request a license from Meta, which Meta may grant at its discretion.

The Llama 3.3 Community License requires anyone who distributes the model or a product containing it to "prominently display “Built with Llama” on a related website, user interface, blogpost, about page, or product documentation."

This is a summary, not the full list. The acceptable use policy applies to internal use too, and if you plan to ship a product or a fine-tuned model, read the license itself.

Which Llama 70B is which?

Three Llama 3 releases carry the 70B name, and search results mix them up. Llama 3 70B had an 8K-token window. Llama 3.1 70B raised it to 128K, and Llama 3.3 70B kept that size and window with new post-training.⁶

Llama 3 70B's window was 8K, so hardware advice written for it may not allow for a 128K conversation. At 128K, the conversation can need as much memory as the 4-bit weights, so check which version and context length a page assumes before you size a machine.

Meta's newer Llama 4 models are a different design: mixture-of-experts models with image input, and none of them is a 70B.

What is Llama 3.3 70B good at?

Meta reports the model against its own earlier models. Its biggest gains over Llama 3.1 70B are in instruction following, math and code, and on several tests it matches or passes the much larger Llama 3.1 405B.

These are Meta's results for the instruction-tuned models, with tests and settings Meta chose; they are not Maple measurements. Meta does not compare Llama 3.3 with models released after it. Newer models report much higher scores on hard reasoning tests; our Gemma 4 31B guide lists Google's results for one of them. Settings differ between makers, so test on your own tasks.

Read the table as a profile, not a ranking. Its highest score in Meta's table is on instruction following, 92.1 on IFEval. In our view, that makes it a fit for:

• drafting, rewriting and editing to a brief
• summarizing and pulling structured fields out of documents
• translation and writing among its eight languages
• tool calling in pipelines already built around the Llama format

How much memory does Llama 3.3 70B need?

Two main amounts, and hardware advice often merges them. The first is the weights, which depend on the precision you load. The second is the KV cache, the memory that holds the conversation and grows with every token in the context. The serving engine's own buffers come on top.

We estimated the second amount from Meta's published architecture.⁸ Llama 3.3 70B uses grouped-query attention: 8 key-value heads of 128 dimensions in each of its 80 layers. At 16-bit precision that comes to 320 KiB per token, about 0.33 MB.

The totals are for one conversation at 16-bit cache precision. Add the serving engine's own overhead. An 8-bit cache roughly halves the KV column, and every extra conversation served at once adds its own.

A full 128K-token conversation needs about as much memory as the 4-bit weights themselves. The 43 GB 4-bit build doesn't fit entirely on a single 24 GB or 32 GB consumer GPU. Runtimes such as Ollama can split the model between GPU and system memory, but expect it to run much slower. As rough budgets by our estimate, two 24 GB cards can just hold the 4-bit build with a short context, and an 80 GB GPU holds 4-bit weights with about 64K tokens of context.

How do you run Llama 3.3 70B yourself?

Fastest on several GPUs or one large one. Running the weights yourself gives you the most control over the model, the serving stack and where your data goes. The cost is hardware, setup and keeping the stack current.

Meta's own downloads, on its site and on Hugging Face, ask you to accept the license first; converted builds such as Ollama's carry the same license. Then pick a runtime:

• Ollama. ollama run llama3.3 pulls the 43 GB 4-bit build, and other tags range from 26 GB at 2-bit to 141 GB at 16-bit. Lower-bit builds save memory at some cost in quality, so test them on your own prompts.
• vLLM. The vLLM project's recipe lists at least 170 GB of GPU memory for BF16, 84 GB for NVIDIA's FP8 checkpoint and 42 GB for NVIDIA's 4-bit NVFP4 checkpoint on Blackwell GPUs.⁹ Set --max-model-len to the context you need, since the full 128K costs memory you may want elsewhere.

Meta's reference code defaults to temperature 0.6 and top-p 0.9, a sensible starting point. Ask whether the operational work buys you something a hosted route can't.

Self-hosting is the right call when:

• your information must never leave hardware you control
• you need to fine-tune the model, which the license allows
• you already have about 48 GB of GPU memory and only need a short context

If you have one 80 GB GPU, a mixture-of-experts model may serve you better. Our GPT-OSS 120B guide covers a larger model whose 4-bit weights fit on one.

How do you use Llama 3.3 70B in Maple?

Choose Llama 3.3 70B under More models in the model picker in Maple Chat. It is on every plan, including Free. We configure it with the full 131,072-token window and Meta's recommended sampling settings.

There's no thinking control to set, because the model has no reasoning mode. Its knowledge stops at the end of 2023, so turn on web search in the composer for anything more recent.

For developer tools and existing applications, Maple Proxy is our API route, with the model ID llama3-3-70b. It supports tool calling there, and API access comes with Pro, Max and Team. The Proxy model list shows its current context window and features.

How it counts against your plan depends on the plan:

• On Free, your daily allowance is the same whichever model you choose.
• On Pro, Max and Team, chat and API usage draw from your plan at each model's per-token rate, listed on our pricing page.

Llama 3.3 70B's per-token rate is several times that of GPT-OSS 120B, and higher than Gemma 4 31B's, GLM-5.3 Flash's and DeepSeek V4.1 Flash's. For the same number of tokens, it uses up a paid plan faster than any of those. The ratio to GPT-OSS 120B is largest for input tokens, so long pasted documents show the difference most.

The route has limits to plan around:

• Text only. On paid plans, a vision model describes an uploaded image first, and Llama 3.3 70B answers from that description.¹⁰ The Free plan doesn't include image upload.
• No reasoning mode. For multi-step math, code or analysis, pick a reasoning model.
• Higher per-token rate. On paid plans, the lower-priced models above cost less for the same tokens.

When is another Maple model a better fit?

Often. The newer models in our picker add reasoning, image input or longer windows. This table matches common needs to the model we would try first, linked to its guide.

Llama 3.3 70B is still a sound choice for text work to a clear brief, when your prompts were written and tested for Llama, or when you want a second model on the Free plan beside GPT-OSS 120B. Every model we offer sits in our model library.

What information reaches each route?

That depends on who operates the model: you, a hosting service, or Maple. Three questions get blurred together. Who can read your prompts while they are processed, what happens to them afterward, and is any of it used for training?

Self-hosted weights. With a fully local setup and no outside tools, your prompts and outputs stay on infrastructure you operate. The trust boundary is your own access controls, logging and retention. That makes this the strongest option when information must stay on your hardware, and it makes you responsible for securing it.

A hosting service. Meta publishes the weights, and many services run Llama 3.3 70B behind an API. Each sets its own terms. Read them for the service and plan you would use: whether prompts are retained, who can access them, and whether they are used for training.

Maple. We run Meta's open weights inside an attested enclave. Your chats are encrypted before they leave your device, and our proof page explains the attestation you can check yourself.

The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt. It works like certificate chains on the web: you check Maple's enclave, and Maple's enclave checks the model's. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the quality of the answer. A private route doesn't make the model's output more accurate, so review its facts, code and suggestions as you would any untrusted input.

Which route fits your task?

It depends on where your information may go, how much hardware you can run and how long your context is. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware, you want to fine-tune, or you already have the GPU memory.
• Choose a hosting service if your prompts aren't sensitive and its terms fit your data.
• Choose Maple if you want Llama 3.3 70B without 48 GB of GPUs and without a host that can read your chats.

Sources

1. Meta, Llama 3.3 model card, llama-models repository, last updated February 13, 2025. Source for the release date, parameters, context length, input and output types, supported languages, training data and cutoff, fine-tuning method and the benchmark results; Meta's Hugging Face listing is the source for the exact parameter count and the Llama 3.1 70B base.
2. Meta, Llama 3.3 Community License Agreement, December 6, 2024, and the Llama 3.3 Acceptable Use Policy. Source for the license conditions and the quotation above.
3. Hugging Face, repository file listing for meta-llama/Llama-3.3-70B-Instruct, our sum of the 30 safetensors files: 141.1 GB at BF16; and Ollama, llama3.3 tags, the source for the 4-bit, 8-bit, 2-bit and 16-bit download sizes and the default tag.
4. Maple, public model configuration, its sampling settings and the app model picker. Source for the Free plan tier, the configured window, text-only input, no reasoning and the sampling settings in Maple.
5. Meta, Llama 3.3 prompt formats. Source for zero-shot function calling, JSON function definitions, the list-of-calls reply format and parallel calls.
6. Meta, model cards for Llama 3, Llama 3.1 and Llama 4. Source for the release dates, context lengths and the Llama 4 design.
7. Hugging Face, repository file listing for NVIDIA's Llama 3.3 70B Instruct FP8. Our sum of the 15 safetensors files: 72.7 GB.
8. Maple's calculation from Meta's architecture parameters for Llama 3.3 70B: 80 layers, 64 attention heads of 128 dimensions and 8 key-value heads, with keys and values at 2 bytes each. Meta's reference generation code is the source for the default temperature and top-p.
9. vLLM project, Llama 3.3 70B serving recipe, updated September 22, 2026. Source for the minimum GPU memory at each precision and the NVFP4 hardware.
10. Maple, image description in the app, the step that describes images before the chosen model answers and its paid-plan check. Source for how the app handles an uploaded image.