How much memory do I need to run an LLM locally?
Multiply the parameter count in billions by the bits per weight and divide by 8 to get gigabytes, then add 10 to 25% for context and the system. At 4-bit, a 70B model needs about 40 GB, a 235B model about 130 GB, and a 671B model such as DeepSeek V3.1 about 380 GB.
Mixture-of-experts (MoE) models still need memory for all their parameters, but only a fraction are used for each token, so they run much faster than their size suggests.
Context needs memory too. Long conversations and many simultaneous users can add tens of gigabytes on top of the weights.
On a Mac, the GPU can use only part of system memory by default: about two thirds on smaller machines and three quarters on larger ones. The limit can be raised with the iogpu.wired_limit_mb setting. Multi-Mac rows plan on about 160 GB of weights per 256 GB Mac Studio.
| Model | Parameters (total / active) | Weights | Runs on, with headroom |
|---|---|---|---|
| gpt-oss-20b | 21B / 3.6B | 14 GB (MXFP4) | 24 GB laptop |
| Qwen3.8-27B | 27B dense | 16 GB (4-bit), 30 GB (8-bit) | 32 GB laptop |
| Llama 3.3 70B | 70B dense | 40 GB (4-bit), 75 GB (8-bit) | 64–128 GB Mac |
| gpt-oss-120b | 117B / 5.1B | 65 GB (MXFP4) | 96–128 GB Mac |
| Qwen3-235B-A22B | 235B / 22B | 132 GB (4-bit), 250 GB (8-bit) | 256 GB Mac Studio |
| DeepSeek V3.1 | 671B / 37B | 378 GB (4-bit), 713 GB (8-bit) | 3 Mac Studios (4-bit) |
| Kimi K2 Thinking | 1T / 32B | 594 GB (native 4-bit) | 4 Mac Studios |
| DeepSeek V4 Pro | 1.6T / 49B | 893 GB (as released) | 6 Mac Studios |
Sources
- local.ai: which models fit your hardware
- Hugging Face: openai/gpt-oss-120b
- Hugging Face: Qwen/Qwen3.8-27B
- Hugging Face: meta-llama/Llama-3.3-70B-Instruct
- Hugging Face: Qwen/Qwen3-235B-A22B-Instruct-2507
- Hugging Face: deepseek-ai/DeepSeek-V3.1
- Hugging Face: moonshotai/Kimi-K2-Thinking
- Hugging Face: deepseek-ai/DeepSeek-V4-Pro-0813
Updated