Best LLMs for 4GB VRAM
With 4 GB of VRAM you can run models up to Gemma 4 E2B (5.1B) at Q4_K_M with an 8K context. 7 of 58 tracked models fit fully, and 2 more fit at a lower quantization.
Models that fit in 4 GB
| Model | VRAM at 8K | Best quant that fits | Max context at Q4 |
|---|---|---|---|
| Gemma 4 E2B | 4.0 GB | IQ4_XS | – |
| Gemma 3 4B | 3.7 GB | Q4_K_M | 9K |
| Qwen 3.5 4B | 3.7 GB | Q4_K_M | 10K |
| Ministral 3 3B | 3.5 GB | Q4_K_M | 10K |
| Llama 3.2 3B | 3.5 GB | Q4_K_M | 10K |
| Llama 3.2 1B | 1.7 GB | FP16 / BF16 | 73K |
| Gemma 3 1B | 1.6 GB | FP16 / BF16 | 32K |
Fits at a lower quantization
These need more than 4 GB at Q4_K_M but fit at a smaller format, with more quality loss.
| Model | VRAM at Q4_K_M | Fits at |
|---|---|---|
| Qwen 3 4B | 4.2 GB | Q2_K |
| Phi-4 Mini | 4.1 GB | Q3_K_M |
Image and video models
Best weight precision that fits 4 GB with model swapping.
| Model | Best precision |
|---|---|
| Stable Diffusion 1.5 | FP16 / BF16 |
| SDXL 1.0 | Needs offloading or more VRAM |
| Stable Diffusion 3.5 Medium | Needs offloading or more VRAM |
| Stable Diffusion 3.5 Large | Needs offloading or more VRAM |
| FLUX.1 [dev] | Needs offloading or more VRAM |
| FLUX.1 [schnell] | Needs offloading or more VRAM |
| Qwen-Image | Needs offloading or more VRAM |
| Wan 2.2 A14B (T2V / I2V) | Needs offloading or more VRAM |
| Wan 2.2 TI2V 5B | Needs offloading or more VRAM |
| HunyuanVideo | Needs offloading or more VRAM |
| FLUX.1 Kontext [dev] | Needs offloading or more VRAM |
| FLUX.2 [dev] | Needs offloading or more VRAM |
| Z-Image Turbo | Needs offloading or more VRAM |
| Qwen-Image-Edit 2509 | Needs offloading or more VRAM |
| LTX-2 | Needs offloading or more VRAM |
4 GB VRAM: common questions
Is 4 GB of VRAM enough for local LLMs?
Yes for models up to about 5.1B at Q4_K_M with 8K context. 4 GB fits 7 of the 58 models tracked here, including Gemma 4 E2B, Gemma 3 4B, Qwen 3.5 4B.
What is the biggest model 4 GB can run?
Gemma 4 E2B (5.1B) at Q4_K_M with 8K context, using about 4.0 GB. Qwen 3 4B also fits at Q2_K, with some quality loss.
Is 4 GB enough for Flux and SDXL?
SDXL: Needs offloading or more VRAM. FLUX.1 dev: Needs offloading or more VRAM (best precision that fits with model swapping).