Models · chat and writing, reasoning and maths
Llama 3.3 70B: VRAM requirements and GPUs that run it
Meta's big dense model. Needs 48 GB at Q4 or a well-stocked Mac. At Q4 with 8K context it needs about 48 GB, so you want a GPU with at least 64 GB of VRAM or a Mac with 64 GB of unified memory.
Memory needed by quantization
| Quant | Weights | Total at 8K | Total at 32K |
|---|---|---|---|
| Q4_K_M | 46 GB | 48 GB | 56 GB |
| Q5_K_M | 55 GB | 58 GB | 65 GB |
| Q8_0 | 82 GB | 85 GB | 93 GB |
| FP16 | 155 GB | 158 GB | 166 GB |
Totals include about 10% runtime overhead and the KV cache. Longer conversations cost real memory: going from 8K to 32K context adds about 7.7 GB at any quantization. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.
Which GPUs can run Llama 3.3 70B?
No single consumer GPU we track runs it fully in VRAM at 8K context. 4 more can run it with part of the model in system RAM, which is noticeably slower.
| GPU | VRAM | Status | Quant | Speed |
|---|---|---|---|---|
| RTX 5090 | 32 GB | Offload | Q3_K_M | ~4.0 tok/s |
| RTX 4090 | 24 GB | Offload | Q3_K_M | ~2.0 tok/s |
| RTX 3090 / 3090 Ti | 24 GB | Offload | Q3_K_M | ~1.9 tok/s |
| RX 7900 XTX | 24 GB | Offload | Q3_K_M | ~2.0 tok/s |
Running it on a Mac
Apple Silicon Macs can give a model about 72% of their unified memory. Llama 3.3 70B fits from 64 GB (Apple M4 Pro, about 5.2 tokens per second). With a faster chip like the Apple M2 Ultra (64 GB) expect around 15 tokens per second. Compare the options in our Mac vs NVIDIA guide.
Similar size models: DeepSeek-R1 Distill 70B.
Compare Llama 3.3 70B
Questions
How much VRAM does Llama 3.3 70B need?
About 48 GB at Q4_K_M with 8K context, 85 GB at Q8_0, and 158 GB at full FP16. In practice that means a card with at least 64 GB of VRAM for Q4.
Can Llama 3.3 70B run on an 8 GB GPU?
No. It needs about 48 GB at Q4, well beyond 8 GB even with offload.
Can I run Llama 3.3 70B on a Mac?
Yes. The smallest configuration that runs it fully is Apple M4 Pro with 64 GB of unified memory, at about 5.2 tokens per second. More memory bandwidth (Pro, Max, Ultra chips) means faster replies.
Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.