Local LLM Hardware Stats
How much memory each model tier needs, and what fits your device. Derived from ModelFit's 129 local models. Updated 2026-10-26.
- Local models
- 129
- Model families
- 32
- 16GB sweet spot
- ~14B
- 64GB+ ceiling
- ~70B
How much RAM do you need to run a local LLM? An 8GB device runs models up to ~9B, a 16GB device comfortably runs up to ~14B, 32GB unlocks ~35B-class models, and 64GB+ runs 70B-class models. At Q4_K_M a model needs roughly 0.6 GB per billion parameters, and ModelFit budgets ~70% of unified memory for the model on machines up to 32GB, scaling to ~85% at 128GB and above. High-RAM Macs can wire even more memory to the GPU by raising iogpu.wired_limit_mb. ModelFit tracks 129 local models across 32 families.
Model size by memory budget
| Memory (RAM / VRAM) | Model budget (~70-85%) | Runs up to (dense) | Models that fit | Strong pick |
|---|---|---|---|---|
| 8 GB | ~5.6 GB | ~9B params | 33 / 129 | Ornith 1.0 9B |
| 12 GB | ~8.4 GB | ~12B params | 50 / 129 | Gemma 4 12B |
| 16 GB | ~11.2 GB | ~14B params | 59 / 129 | Qwen3.5 9B Instruct (Q8) |
| 24 GB | ~16.8 GB | ~29.3B params | 73 / 129 | Qwen3.8 27B |
| 32 GB | ~22.4 GB | ~35B params | 90 / 129 | Qwen3.6 35B-A3B |
| 36 GB | ~25.4 GB | ~35B params | 95 / 129 | Qwen3.6 35B-A3B |
| 48 GB | ~34.8 GB | ~35B params | 102 / 129 | Qwen3.8 27B (Q8) |
| 64 GB | ~48 GB | ~70B params | 108 / 129 | Qwen3.6 35B-A3B (Q8) |
| 72 GB | ~54.9 GB | ~70B params | 110 / 129 | Qwen3.6 35B-A3B (Q8) |
| 96 GB | ~76.8 GB | ~123B params | 118 / 129 | Qwen3.5 122B-A10B Instruct |
| 128 GB | ~108.8 GB | ~123B params | 124 / 129 | Qwen3.5 122B-A10B Instruct |
| 192 GB | ~163.2 GB | ~123B params | 126 / 129 | Qwen3 235B A22B |
| 256 GB | ~217.6 GB | ~123B params | 126 / 129 | Qwen3 235B A22B |
| 512 GB | ~435.2 GB | ~405B params | 129 / 129 | Qwen3 235B A22B |
Q4_K_M assumed. tok/s and fit are estimates from ModelFit's dataset, not measured benchmarks.
Key facts
- ModelFit tracks 172 AI models across 32 families; 129 run locally via Ollama on Apple Silicon or NVIDIA GPUs (ModelFit, 2026).
- At Q4_K_M quantization, a local LLM needs roughly 0.6 GB of memory per billion parameters for the weights, rising to about 0.8 GB per billion for models under 8B once runtime overhead is counted (ModelFit, 2026).
- ModelFit sizes recommendations to ~70% of a device’s unified memory up to 32GB, scaling to ~85% at 128GB and above, leaving headroom for the OS, context, and KV-cache (ModelFit, 2026).
- An 8GB device comfortably runs dense local models up to ~9B parameters at Q4; 33 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 12GB device comfortably runs dense local models up to ~12B parameters at Q4, or up to ~12.1B total as a mixture-of-experts model; 50 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 16GB device comfortably runs dense local models up to ~14B parameters at Q4; 59 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 24GB device comfortably runs dense local models up to ~29.3B parameters at Q4; 73 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 32GB device comfortably runs dense local models up to ~35B parameters at Q4; 90 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 36GB device comfortably runs dense local models up to ~35B parameters at Q4, or up to ~39.8B total as a mixture-of-experts model; 95 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 48GB device comfortably runs dense local models up to ~35B parameters at Q4, or up to ~46.7B total as a mixture-of-experts model; 102 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 64GB device comfortably runs dense local models up to ~70B parameters at Q4, or up to ~78.1B total as a mixture-of-experts model; 108 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 72GB device comfortably runs dense local models up to ~70B parameters at Q4, or up to ~80B total as a mixture-of-experts model; 110 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 96GB device comfortably runs dense local models up to ~123B parameters at Q4, or up to ~124B total as a mixture-of-experts model; 118 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 128GB device comfortably runs dense local models up to ~123B parameters at Q4, or up to ~125B total as a mixture-of-experts model; 124 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 192GB device comfortably runs dense local models up to ~123B parameters at Q4, or up to ~235B total as a mixture-of-experts model; 126 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 256GB device comfortably runs dense local models up to ~123B parameters at Q4, or up to ~235B total as a mixture-of-experts model; 126 of ModelFit’s 129 local models fit (ModelFit, 2026).
- A 512GB device comfortably runs dense local models up to ~405B parameters at Q4, or up to ~671B total as a mixture-of-experts model; 129 of ModelFit’s 129 local models fit (ModelFit, 2026).
Frequently asked questions
How much RAM do I need to run a local LLM?
8GB runs models up to ~9B, 16GB comfortably runs up to ~14B, 32GB unlocks ~35B-class models, and 64GB or more runs 70B-class models. ModelFit sizes this at ~70% of unified memory up to 32GB, scaling to ~85% at 128GB+, since a Q4 model needs roughly 0.6 GB per billion parameters.
What size LLM can I run on 16GB of RAM?
On 16GB you can comfortably run local models up to ~14B parameters at Q4; 59 of ModelFit's 129 local models fit, with Qwen3.5 9B Instruct (Q8) a strong pick.
Can 8GB of RAM run a local LLM?
Yes. 8GB runs 33 of ModelFit's local models: small 3-4B models up to ~9B at Q4. Expect tight headroom; close other apps for the best speed.
Cite this page
Free to reuse with attribution (CC BY 4.0). Full machine-readable data: the compatibility dataset and a JSON export.
ModelFit: Local LLM Hardware Stats (2026). https://modelfit.io/stats/ (dataset updated 2026-10-26).