Llama 3.3 70B GPU requirements: 35.3 GB at 4-bit, 141 GB at BF16
Llama 3.3 70B needs 141 GB of GPU memory for its weights at BF16, 70.6 GB at FP8 and 35.3 GB at 4-bit, plus 10.7 GB of KV cache for a 32K context. The cheapest node that runs it here is 1× A100 PCIe 80GB at 4-bit, for $0.09 an hour on spot.
Weights plus a 32K 16-bit KV cache, with 10% of the node's memory kept free. Billed per minute.
- Parameters
- 70.6BDense
- Weights, 4-bit
- 35.3 GBBF16: 141 GB
- KV cache, 32K
- 10.7 GBOne sequence, 16-bit cache
- Context window
- 128KNative
How much VRAM Llama 3.3 70B needs, format by format.
Weights are parameters × bytes per parameter. Each context column adds the KV cache of one sequence of that length.
| Format | Weights | + 8Kcontext | + 32Kcontext | + 128Kcontext | Cheapest node, 32K |
|---|---|---|---|---|---|
| BF16Published weights, fine-tuning | 141 GB | 144 GB | 152 GB | 184 GB | 4× A40 · $0.24/h |
| FP8Hopper, Ada, Blackwell, MI300 | 70.6 GB | 73.2 GB | 81.3 GB | 113 GB | 2× L40 · $0.18/h |
| 4-bitAWQ, GPTQ, GGUF Q4 | 35.3 GB | 38.0 GB | 46.0 GB | 78.2 GB | 1× A100 PCIe 80GB · $0.09/h |
The serving engine adds its own buffers, and an 8-bit KV cache halves the context columns. We keep 10% of a node's memory free when we say a model fits. Speed is a separate question: it follows memory bandwidth and, for a mixture of experts, the active parameters.
GPUs that run Llama 3.3 70B, cheapest first.
For each format, the smallest node of each GPU model that holds the weights and a 32K context, sorted by spot price. One row per GPU model; larger nodes of the same GPU work too.
| Format | GPU | GPU memory | Spot | On-demand | 730 h spot | Action |
|---|---|---|---|---|---|---|
| 4-bit | 1× A100 PCIe 80GBAmpere · 46.0 GB needed | 80 GB | $0.09 | $0.19 | $65.70 | Deploy |
| 4-bit | 4× RTX A4000Ampere · 46.0 GB needed | 64 GB | $0.12 | $0.20 | $87.60 | Deploy |
| 4-bit | 2× A40Ampere · 46.0 GB needed | 96 GB | $0.12 | $0.18 | $87.60 | Deploy |
| 4-bit | 4× T4Turing · 46.0 GB needed | 64 GB | $0.16 | $0.28 | $116.80 | Deploy |
| FP8 | 2× L40Ada Lovelace · 81.3 GB needed | 96 GB | $0.18 | $0.38 | $131.40 | Deploy |
| FP8 | 2× RTX 6000 AdaAda Lovelace · 81.3 GB needed | 96 GB | $0.18 | $0.38 | $131.40 | Deploy |
| FP8 | 1× RTX PRO 6000 BlackwellBlackwell · 81.3 GB needed | 96 GB | $0.19 | $0.39 | $138.70 | Deploy |
| FP8 | 4× L4Ada Lovelace · 81.3 GB needed | 96 GB | $0.20 | $0.32 | $146.00 | Deploy |
| BF16 | 4× A40Ampere · 152 GB needed | 192 GB | $0.24 | $0.36 | $175.20 | Deploy |
| BF16 | 4× L40Ada Lovelace · 152 GB needed | 192 GB | $0.36 | $0.76 | $262.80 | Deploy |
| BF16 | 4× RTX 6000 AdaAda Lovelace · 152 GB needed | 192 GB | $0.36 | $0.76 | $262.80 | Deploy |
| BF16 | 4× RTX A6000Ampere · 152 GB needed | 192 GB | $0.36 | $0.68 | $262.80 | Deploy |
Spot capacity can be reclaimed with a 2-minute notice; for a live endpoint, use on-demand. How spot works · serving a language model.
What a long context costs in memory.
Grouped-query attention, 8 KV heads of 128 across 80 layers.
| Context | KV cache, 1 sequence | 8 sequences |
|---|---|---|
| 8K tokens | 2.68 GB | 21.5 GB |
| 32K tokens | 10.7 GB | 85.9 GB |
| 128K tokens | 42.9 GB | 344 GB |
Serving several users
KV cacheEvery concurrent request keeps its own cache, so memory for context grows with users × context length. A server batching 8 requests of 32K needs eight times the single-sequence figure — 85.9 GB here.
- Recommended node1× A100 PCIe 80GB
- Weights at 4-bit35.3 GB
- Context that fits, 1 sequence109K
Paged attention (vLLM, SGLang) allocates the cache in blocks as requests grow, so a node holds more short requests than the worst case suggests.
Llama 3.3 70B, as published.
- Developer
- MetaThe common 70B baseline (same size and layout as Llama 3.1 70B)
- Parameters
- 70.6B
- Architecture
- Dense model, 80 layersGrouped-query attention, 8 KV heads of 128 across 80 layers
- Context window
- 131,072 tokensAs set in the published configuration
- Published format
- BF16, 141 GB
- Licence
- Llama 3.3 Community License
- Released
- 6 December 2024
- Source
- meta-llama/Llama-3.3-70B-Instructconfig.json read 26 October 2026
Serve it with vLLM
template: vLLMOn 1× A100 PCIe 80GB, from the instance terminal:
$ vllm serve meta-llama/Llama-3.3-70B-Instruct \ --tensor-parallel-size 1 \ --max-model-len 32768
Very recent architectures need a recent vLLM; the model card gives the minimum version and any extra flags.
Llama 3.3 70B on a GPU — the questions that come up
How much VRAM does Llama 3.3 70B need?
For the weights alone: 141 GB at BF16, 70.6 GB at FP8 and 35.3 GB at 4-bit. Add 10.7 GB of KV cache for one 32K-token context, or 42.9 GB at the full 128K window, and keep about 10% of the memory free for activations.
Can Llama 3.3 70B run on a single GPU?
One RTX 4090 (24 GB): no. One RTX 5090 (32 GB): no. One H100 SXM (80 GB): yes, at 4-bit. One H200 SXM (141 GB): yes, at FP8. One B200 SXM (180 GB): yes, at BF16. Those answers assume a 32K context and the widest format that still fits.
What is the cheapest way to run Llama 3.3 70B?
Here, 1× A100 PCIe 80GB at 4-bit: $0.09 per hour on spot and $0.19 on-demand, billed per minute — $65.70 for a 730-hour month on spot. That node has 80 GB of GPU memory for the 46.0 GB this setup needs.
How long a context fits on the recommended node?
With the weights at 4-bit on 1× A100 PCIe 80GB, the remaining memory holds a 16-bit KV cache for about 112,075 tokens of context for one sequence (109K). An 8-bit KV cache doubles that; serving several users at once divides it between them.
Which format should I run it in?
BF16 keeps the published weights untouched and is the format for fine-tuning. FP8 halves the memory and runs in hardware on Hopper, Ada, Blackwell and MI300-class GPUs. 4-bit quantisation (AWQ, GPTQ, GGUF) quarters it, at a quality cost that is larger on small models than on big ones; check it against your own evaluation before relying on it.
How do I deploy it?
Pick the node in the table above and launch it with the vLLM template; the model downloads from Hugging Face (meta-llama/Llama-3.3-70B-Instruct) on first start. Spread the weights over the node with --tensor-parallel-size equal to the GPU count, and set --max-model-len to the context you need. New architectures need a recent vLLM: check the model card for the minimum version.
Sizing another model? GPU requirements for 64 open models · serving a language model · fine-tuning.
Llama 3.3 70B on 1× A100 PCIe 80GB for $0.09 an hour.
Launch the node with the vLLM template, point it at meta-llama/Llama-3.3-70B-Instruct, stop it when you are done. Prepaid, billed by the minute.
Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.