How much VRAM does Llama 3.3 70B need?
Llama 3.3 70B needs 39.8 GB for its weights at 4-bit, 43.3 GB with an 8K context and 50.8 GB with 32K. At 8-bit, 73.3 GB.
- Parameters
- 70.5B
- Longest context
- 128K
- Released
- Dec 2024
- License
- Llama 3.3 Community
- Strongest at
- Chat, Coding, Agents, Docs
Memory at each precision
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| 16-bit | 131 GB | 135 GB | 142 GB | 172 GB |
| 8-bit | 69.8 GB | 73.3 GB | 80.8 GB | 111 GB |
| 6-bit | 53.9 GB | 57.3 GB | 64.8 GB | 94.8 GB |
| 5-bit | 46.7 GB | 50.2 GB | 57.7 GB | 87.7 GB |
| 4-bit | 39.8 GB | 43.3 GB | 50.8 GB | 80.8 GB |
| 3-bit | 32.1 GB | 35.6 GB | 43.1 GB | 73.1 GB |
| 2-bit | 24.6 GB | 28.1 GB | 35.6 GB | 65.6 GB |
| 1.6-bit | 13.6 GB | 17.0 GB | 24.5 GB | 54.5 GB |
One conversation, 16-bit context cache, runtime overhead included.
What runs it
| Machine | Best fit | Context | Speed | First reply |
|---|---|---|---|---|
| RTX 3060 12 GB | Too big | |||
| RTX 4060 Ti 16 GB | Too big | |||
| RTX 3090 | Too big | |||
| RTX 4090 | Too big | |||
| RTX 5090 | Too big | |||
| 2× RTX 3090 | 3-bit | 36K | 18 tok/s | 2.4 s |
| RTX PRO 6000 Blackwell | 8-bit | 64K | 16 tok/s | 1.1 s |
| M4 Max 128 GB | 8-bit | 64K | 4.7 tok/s | 24 s |
| M5 Max 128 GB | 8-bit | 64K | 5.2 tok/s | 6.9 s |
| Ryzen AI Max+ 395 128 GB | 8-bit | 64K | 2.1 tok/s | 14 s |
| DGX Spark 128 GB | 8-bit | 128K | 2.6 tok/s | 4.8 s |
| M3 Ultra 256 GB | 8-bit | 128K | 6.9 tok/s | 13 s |
| M5 Ultra 512 GB | 8-bit | 128K | 9.9 tok/s | 4.2 s |
| H100 SXM | 8-bit | 16K | 30 tok/s | 1.0 s |
| DGX Station GB300 | 8-bit | 128K | 58 tok/s | 0.4 s |
| 8× H200 NVL | 8-bit | 128K | 41 tok/s | 0.2 s |
Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.
How to run it
- Ollama
ollama run llama3.3:70b- llama.cpp
llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF- vLLM
vllm serve meta-llama/Llama-3.3-70B-Instruct- MLX (Mac)
mlx_lm.chat --model mlx-community/Llama-3.3-70B-Instruct-4bit
Questions
How much VRAM does Llama 3.3 70B need?
Llama 3.3 70B needs 39.8 GB for its weights at 4-bit, 43.3 GB with an 8K context and 50.8 GB with 32K. At 8-bit, 73.3 GB.
Can Llama 3.3 70B run on an RTX 4090?
Not on one card. The smallest machine here that runs it is the 2× RTX 3090.