Can I run Llama 3.3 70B Instruct locally?
Llama 3.3 70B Instruct by meta-llama has 71B parameters. At 4-bit with an 8K context it needs about 43.3 GB of GPU memory. It fits on 25 of 65 GPUs and Macs we track. See all models that fit in 48 GB of VRAM.
GGUF quant file sizes were not in size order, so this page uses size estimates instead of real file sizes.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) | rough | 26.5 GB | 27.8 GB | 35.3 GB | 65.3 GB |
| 3-bit (Q3_K_M) | usable | 34.0 GB | 35.3 GB | 42.8 GB | 72.8 GB |
| 4-bit (Q4_K_M) | good, most popular | 42.1 GB | 43.3 GB | 50.8 GB | 80.8 GB |
| 5-bit (Q5_K_M) | very good | 48.7 GB | 50.0 GB | 57.5 GB | 87.5 GB |
| 6-bit (Q6_K) | near lossless | 55.8 GB | 57.0 GB | 64.5 GB | 94.5 GB |
| 8-bit (Q8_0) | effectively lossless | 71.7 GB | 73.0 GB | 80.5 GB | 110 GB |
| 16-bit (F16/BF16) | original weights | 133 GB | 135 GB | 142 GB | 172 GB |
Which GPUs and Macs run Llama 3.3 70B Instruct?
Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).
Prefer an API?
Meta: Llama 3.3 70B Instruct is on OpenRouter at $0.71 per 1M input tokens and $0.71 per 1M output tokens. Price details and history →
Questions
How much VRAM does Llama 3.3 70B Instruct need?
About 43.3 GB at 4-bit (Q4_K_M) with an 8K context: 40.2 GB of weights (estimated), 2.5 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 73.0 GB.
What is the smallest GPU that runs Llama 3.3 70B Instruct?
No single consumer GPU in our table fits it at Q4_K_M with 8K context; the smallest option that does is the H100 PCIe (80 GB).
Can I run Llama 3.3 70B Instruct on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M2 Max 96GB or larger, using the default macOS GPU memory limit.
Similar models
Same family or a similar parameter count, with the VRAM each needs at Q4_K_M and 8K context.
Llama 3.1 70B
71B · 42.7 GB at 4-bitLlama 4 Scout 17B 16E Instruct
109B · 63.2 GB at 4-bitLlama 4 Scout 17B 16E
109B · 64.1 GB at 4-bitDeepSeek R1 Distill Llama 70B
71B · 43.3 GB at 4-bitKimi Dev 72B
73B · 44.6 GB at 4-bitHunyuan A13B Instruct
80B · 47.1 GB at 4-bitModel facts
| Hugging Face repo | meta-llama/Llama-3.3-70B-Instruct |
| Parameters | 70,553,706,496 (71B) |
| Layers / hidden size | 80 / 8192 |
| Attention | GQA; 64 heads, 8 KV heads, head dim 128 |
| KV cache per token (FP16) | 320.0 KiB |
| Max context | 131,072 tokens |
| License | llama3.3 (gated: accept terms on Hugging Face) |
| Released on HF | Nov 26, 2024 |
| Downloads (30d) / likes | 344,323 / 3,208 |
Sources: Hugging Face model API · config.json (unsloth/Llama-3.3-70B-Instruct-GGUF) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.