How much VRAM do you need for Llama 3 and Llama 3.3 70B?

Short answer: to run Llama 3.3 70B (or the older Llama 3 and 3.1 70B) fully on your graphics card in Ollama, you need about 48GB of VRAM. Ollama's default 4-bit download, llama3.3:70b, is 43GB, and the model needs a few more GB for its context window. That means two 24GB cards (two RTX 3090s or two RTX 4090s), one 96GB RTX PRO 6000, or a Mac or mini PC with 64GB or more of unified memory. The 8-bit version is 75GB (plan on about 80GB) and full FP16 is 141GB. A single 24GB or 32GB card can still run it, but part of the model spills into system RAM and replies get much slower.

Every model size below comes from Ollama's own tag pages for llama3.3, llama3.1 and llama3, and every VRAM number comes from NVIDIA's own product pages, checked Nov 3, 2026. No prices here because they change daily. Click through for the current price.

Llama 3 70B vs 3.1 vs 3.3: same memory, different context

All three are 70 billion parameter models from Meta, and Ollama lists the exact same file sizes for each quantization. Llama 3.3 70B is the newest (Meta released it Dec 6, 2024). It is text only, supports a 128K token context window, and Ollama describes it as offering similar performance to the much bigger Llama 3.1 405B. The original Llama 3 70B only handles 8K tokens of context. If you are choosing today, pull llama3.3:70b. The VRAM math is the same for all three.

The quantization math, step by step

A model's memory is mostly its weights: parameters x bits per weight / 8 = bytes. Quantization just means storing each weight with fewer bits.

• FP16 (full quality): 16 bits = 2 bytes per weight. 70 billion x 2 bytes = 140GB. Ollama's fp16 file is 141GB.
• Q8_0 (8-bit): 8 bits per weight plus a small scale value for every block of 32 weights, which works out to about 8.5 bits. 70 billion x 8.5 / 8 = about 74GB. Ollama's q8_0 file is 75GB.
• Q4_K_M (Ollama's default 4-bit): a mix of 4-bit and some higher precision layers. Ollama's file is 43GB, which works out to about 4.9 bits per weight (43GB x 8 / 70 billion).

So going from FP16 to Q4 cuts the memory by about 70%. That is why almost everyone runs 70B models at 4-bit at home.

Every llama3.3:70b download size in Ollama

Download sizes from Ollama's llama3.3 tag page. Bits per weight is our own math (size x 8 / 70 billion). The "fits" column is our guidance and includes room for a normal 4K to 8K context.

Do not forget the context window

On top of the weights, Ollama keeps a memory of the conversation called the KV cache. Meta's Llama 3 paper lists the 70B model with 80 layers, 8 key/value heads and a model dimension of 8,192 split across 64 attention heads (128 per head). At Ollama's default f16 cache that is 2 x 80 x 8 x 128 x 2 bytes = 327,680 bytes, about 0.33MB per token.

Ollama's FAQ says the default context is 4,096 tokens and that you can change it with the OLLAMA_CONTEXT_LENGTH setting or /set parameter num_ctx. It also lets you shrink the cache with OLLAMA_KV_CACHE_TYPE=q8_0, which Ollama says uses about half the memory of f16 with a very small loss in precision. That needs Flash Attention, which Ollama turns on automatically when your GPU supports it. Running more than one chat at once multiplies the context memory, so leave OLLAMA_NUM_PARALLEL at its default of 1 on a home PC.

Quick total for the default setup: 43GB of weights + about 1.3GB of cache at 4K context + a little working room = roughly 45GB to 46GB. That is why 48GB is the practical floor for running it all on the GPU.

Which GPU setups actually fit Llama 3.3 70B

Two cards work because Ollama splits big models. Ollama's FAQ says that if a model does not fit on any single GPU, it is spread across all the available GPUs. You do not need NVLink for this. You do need a motherboard with room for two thick cards and a big power supply: NVIDIA lists 350W of graphics card power and a 750W recommended system for a single RTX 3090, and 450W with an 850W system for a single RTX 4090. For two cards, plan well above that.

One 24GB card plus system RAM is the cheapest way to try it. Ollama fills the GPU first and puts the rest of the model in normal RAM. It works, but the part in system RAM is much slower than VRAM, so expect slow replies. Our guidance: have 64GB of system RAM so Windows and your apps still have room. Run ollama ps and the PROCESSOR column will show the split, for example 48%/52% CPU/GPU.

Unified memory is the other path. Apple silicon Macs and some mini PCs share one pool of memory between CPU and GPU, so a 64GB machine can hold the 43GB file. See our Mac mini M6 vs M5 Pro guide (only the 64GB M5 Pro handles 70B comfortably) and our best mini PC for local AI picks. NVIDIA's DGX Spark has 128GB of unified memory, but NVIDIA lists its memory bandwidth at 273GB/s, far below the 1,792GB/s it lists for the RTX PRO 6000, so expect slower replies than an all-GPU setup.

Is 70B worth it, or should you run a smaller model?

For most people with a 12GB or 16GB card, a smaller model is the better daily driver. llama3:8b is a 4.7GB download in Ollama and runs fast on any card in our best budget GPU for Ollama guide. Step up to 70B when you need better reasoning, writing or coding and you have the memory to run it fully on the GPU. For the full model size chart from 4B to 70B, see how much RAM and VRAM you need for Ollama. Running AI for a whole office? Our small business AI server guide covers multi-user setups. Not sure the hardware is worth it at all? Try the local AI vs cloud AI cost calculator, or start at the Local AI hub.

How to check it on your own PC

1. Install Ollama (our Windows install guide walks through it).
2. Run ollama pull llama3.3:70b. It is a 43GB download, so give it time and disk space.
3. Run ollama run llama3.3:70b, ask a question, then open a second window and run ollama ps. 100% GPU means it fits in VRAM. A CPU/GPU split means part of it is in system RAM.
4. If it splits, try a smaller quant like llama3.3:70b-instruct-q3_K_M (34GB) or a shorter context before you buy anything.

Sources: parameters, context length and release date from Meta's Llama 3.3 70B Instruct model card on Hugging Face. Layer and head counts from Table 3 of Meta's The Llama 3 Herd of Models paper. Download sizes from Ollama's llama3.3, llama3.1 and llama3 tag pages. Context, KV cache and multi-GPU behavior from the Ollama FAQ. VRAM and power from NVIDIA's product pages for the RTX 5090, RTX 4090, RTX 3090, RTX 5060 Ti, RTX PRO 6000 Blackwell and DGX Spark. All checked Nov 3, 2026. "Fits", "slow" and RAM advice are our own guidance, not Meta, Ollama or NVIDIA claims.

FAQ

How much VRAM do I need for Llama 3.3 70B?

About 48GB to run Ollama's default 4-bit version (a 43GB download) fully on the GPU with a normal context window. The 8-bit version needs about 80GB and full FP16 needs about 145GB.

Can I run Llama 3 70B on one RTX 4090 or RTX 3090?

It runs, but not fully on the graphics card. NVIDIA lists 24GB on both cards, which holds about half of the 43GB 4-bit file. Ollama puts the rest in system RAM, so plan on 64GB of RAM and much slower replies.

Can an RTX 5090 run Llama 3.3 70B?

Not fully on the GPU at 4-bit. The RTX 5090 has 32GB and the default 4-bit file is 43GB, so part of it spills into system RAM. Only the 2-bit q2_K file (26GB) fits entirely, and 2-bit is the lowest quality option Ollama offers.

Is Llama 3.3 70B the same size as Llama 3 70B?

In memory, yes. Ollama lists the same sizes for both: 40GB at q4_0, 43GB at q4_K_M, 75GB at q8_0 and 141GB at fp16. The big difference is context: Llama 3 70B handles 8K tokens, Llama 3.1 and 3.3 handle 128K.

Does a longer context window need more VRAM?

Yes. For a 70B Llama model each token of context takes about 0.33MB at Ollama's default f16 cache, so 4K tokens is about 1.3GB, 32K is about 10.7GB and the full 128K is about 43GB. Setting OLLAMA_KV_CACHE_TYPE to q8_0 roughly halves that.