Best local AI models for 4GB VRAM
Ranked by what actually fits at 32K context, computed from real file bytes.
A 4GB card gives you about 3.72 GiB to work with after driver overhead. 17 indexed models fit at 32K context — the largest being s2-pro at 4.6B parameters in Q3_K.
From the file· fit from summed bytesFrom the file· KV per layer
Fits in 4GB at 32K context
largest quantization that fits, per model
| Model | Modality | Best quant | Params○ | Total◐ | Headroom◐ |
|---|---|---|---|---|---|
| Ace-Step1.5 | speech synthesis | F32 | 160M | 2.71 GiB | 1.01 GiB |
| Qwen3-TTS-12Hz-0.6B-Base | speech synthesis | Q4_K | 915M | 1.34 GiB | 2.38 GiB |
| OmniVoice | speech synthesis | F16 | 613M | 2.37 GiB | 1.35 GiB |
| VieNeu-TTS-0.3B | speech synthesis | Q8_0 | 244M | 1.94 GiB | 1.78 GiB |
| neutts-air | speech synthesis | Q8_0 | 748M | 1.90 GiB | 1.82 GiB |
| s2-proMoE | speech synthesis | Q3_K | 4.6B | 3.64 GiB | 0.08 GiB |
| Fun-CosyVoice3-0.5B-2512 | speech synthesis | Q4_K | — | 1.70 GiB | 2.02 GiB |
| VoxCPM2 | speech synthesis | Q8_0 | 2.3B | 3.48 GiB | 0.24 GiB |
| Kokoro-82M | speech synthesis | F16 | 82M | 1.00 GiB | 2.72 GiB |
| VieNeu-TTS | speech synthesis | Q4_0 | 553M | 1.54 GiB | 2.18 GiB |
| csm-1b | speech synthesis | Q8_0 | 1.6B | 3.67 GiB | 0.05 GiB |
| VibeVoice-Realtime-0.5B | speech synthesis | F16 | 1.0B | 2.74 GiB | 0.98 GiB |
| Qwen3-TTS-12Hz-1.7B-Base | speech synthesis | Q8_0 | 1.9B | 2.77 GiB | 0.95 GiB |
| Qwen3-TTS-12Hz-1.7B-VoiceDesign | speech synthesis | Q8_0 | 1.9B | 2.75 GiB | 0.97 GiB |
| VibeVoice-1.5B | speech synthesis | Q4_K | 2.7B | 2.61 GiB | 1.11 GiB |
| VoxCPM-0.5B | speech synthesis | Q4_K | 728M | 3.29 GiB | 0.43 GiB |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | speech synthesis | Q8_0 | 1.9B | 2.75 GiB | 0.97 GiB |
This page models a generic 4GB accelerator, so it answers what fits rather than how fast it runs. For tokens per second you need a specific card — pick one from hardware, where bandwidth is known.