RTX 6000 Ada Generation
NVIDIA
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
Model · Meta
Llama 3.3 70B Instruct: VRAM and hardware requirements
At Q4_K_M with an 8K context, Llama 3.3 70B Instruct needs about 45.8 GB: a GPU with 48 GB or more runs it entirely in VRAM.
- Parameters70.6B
- Max context128K
- LicenseLlama 3.3 Community License
Specifications
- Organization
- Meta
- Architecture
- llama
- Released
- December 2024
- Layers
- 80
- Hidden size
- 8192
- Attention heads
- 64
- KV heads
- 8
- Head dimension
- 128
- Vocabulary
- 128,256
- KV cache per 1K tokens (f16)
- 0.313 GB
- Repository
- meta-llama/Llama-3.3-70B-Instruct
Official specData updated 24 days agoSource: Hugging Face config.json + model card
Quantizations and memory
| Quantization | File size | Memory needed | Minimum GPU memory | Quality |
|---|---|---|---|---|
| BF16 | 131 GB | ~142 GB | – | Reference (unquantized) |
| Q8_0 | 69.8 GB | ~76.8 GB | 80 GB | Near-lossless |
| Q6_K | 53.9 GB | ~60.2 GB | 80 GB | Near-lossless |
| Q5_K_M | 46.5 GB | ~52.7 GB | 80 GB | Very good |
| Q4_K_MMost used | 39.6 GB | ~45.8 GB | 48 GB | Good — the usual choice |
| IQ4_XS | 35.3 GB | ~42 GB | 48 GB | Good — the usual choice |
| Q3_K_M | 31.9 GB | ~38 GB | 48 GB | Noticeable loss |
| Q2_K | 24.6 GB | ~30.8 GB | 32 GB | Significant loss |
File sizes: not available until imported from the model repository. Memory = weights (estimated from bits per weight) + KV cache + runtime overhead + safety margin. (BF16 ≈ 16 bpw, Q8_0 ≈ 8.5 bpw, Q6_K ≈ 6.57 bpw, Q5_K_M ≈ 5.7 bpw, Q4_K_M ≈ 4.9 bpw, IQ4_XS ≈ 4.46 bpw, Q3_K_M ≈ 4 bpw, Q2_K ≈ 3.17 bpw)
Q4, Q5 or Q8: how to choose a quantization
GPUs that run Llama 3.3 70B Instruct
| RTX 509032 GB · 64 GB RAM~45.1 GB · 1.4–2.5 tok/s estimate | Possible but slow |
|---|---|
| RTX 409024 GB · 64 GB RAM~45.1 GB · 0.9–1.7 tok/s estimate | Not recommended |
| RTX 309024 GB · 64 GB RAM~45.1 GB · 0.9–1.7 tok/s estimate | Not recommended |
| RX 7900 XTX24 GB · 64 GB RAM~45.1 GB · 0.9–1.7 tok/s estimate | Not recommended |
| RTX 5060 Ti 16 GB16 GB · 64 GB RAM~45.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| RTX 5070 Ti16 GB · 64 GB RAM~45.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| RTX 508016 GB · 64 GB RAM~45.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| RTX 4060 Ti 16 GB16 GB · 64 GB RAM~45.1 GB · 0.6–1.2 tok/s estimate | Not recommended |
| RTX 4070 Ti SUPER16 GB · 64 GB RAM~45.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| RX 9070 XT16 GB · 64 GB RAM~45.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| RTX 3060 12 GB12 GB · 64 GB RAM~45.1 GB · 0.6–1.1 tok/s estimate | Not recommended |
| RTX 407012 GB · 64 GB RAM~45.1 GB · 0.6–1.1 tok/s estimate | Not recommended |
| RTX 507012 GB · 64 GB RAM~45.1 GB · 0.6–1.2 tok/s estimate | Not recommended |
| RTX 40608 GB · 64 GB RAM~45.1 GB · 0.5–1 tok/s estimate | Not recommended |
Macs that run Llama 3.3 70B Instruct
| M3 Ultra (80-core GPU) · 512 GBMac Studio~44.9 GB · 12–17 tok/s estimate | Excellent |
|---|---|
| M4 Max (40-core GPU) · 128 GBMacBook Pro 14, MacBook Pro 16~44.9 GB · 8–12 tok/s estimate | Compatible |
| M3 Max (40-core GPU) · 128 GBMacBook Pro 14, MacBook Pro 16~44.9 GB · 6–8.5 tok/s estimate | Compatible |
| M4 Pro (20-core GPU) · 64 GBMac mini, MacBook Pro 14~44.9 GB · 4–5.5 tok/s estimate | Compatible |
| M4 Max (32-core GPU) · 36 GBMacBook Pro 14, MacBook Pro 16~44.9 GB · No data | Incompatible |
| M4 · 32 GBMacBook Air, MacBook Pro 14~44.9 GB · No data | Incompatible |
| M5 · 32 GBMacBook Pro 14~44.9 GB · No data | Incompatible |
| M3 · 24 GBMacBook Air, MacBook Pro 14~44.9 GB · No data | Incompatible |
EstimateData updated 24 days agoSource: Run LLM Locally engine
Limits in plain words
The KV cache grows with context: at 32K it adds about 10 GB on top of the weights.
With under 8 GB of VRAM, use a smaller quantization or a smaller model of the same family.
Where to buy
Prices and links follow the market you choose.
Radeon PRO W7900
AMD
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
AMD Ryzen AI Max+ 395
Unified memory PC
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
No current offer imported for this market: the link opens a marketplace search.
Frequently asked questions
How much VRAM does Llama 3.3 70B Instruct need?
About 45.8 GB at Q4_K_M with an 8K context, including the KV cache and runtime overhead. At Q8_0 it needs about 76.8 GB. A longer context adds more: about 10 GB of KV cache at 32K.
Can Llama 3.3 70B Instruct run on a Mac?
Yes on a Mac whose GPU can use at least 44.9 GB, e.g. M4 Pro (20-core GPU) 64 GB. macOS lets the GPU use about two thirds to three quarters of unified memory by default.
How fast is Llama 3.3 70B Instruct?
On a RTX 4090, we estimate 0.9–1.7 tok/s at Q4_K_M. This is an estimate from memory bandwidth (confidence low) until benchmarks are imported.