Best local AI models for 8GB VRAM
Ranked by what actually fits at 32K context, computed from real file bytes.
A 8GB card gives you about 7.44 GiB to work with after driver overhead. 20 indexed models fit at 32K context — the largest being s2-pro at 4.6B parameters in Q8_0.
From the file· fit from summed bytesFrom the file· KV per layer
Fits in 8GB 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 | 4.73 GiB |
| Qwen3-TTS-12Hz-0.6B-Base | speech synthesis | Q4_K_M | 915M | 5.57 GiB | 1.87 GiB |
| OmniVoice | speech synthesis | F32 | 613M | 3.82 GiB | 3.62 GiB |
| orpheus-3b-0.1-ft | speech synthesis | Q6_K | 3.8B | 6.84 GiB | 0.60 GiB |
| VieNeu-TTS-0.3B | speech synthesis | Q8_0 | 244M | 1.94 GiB | 5.50 GiB |
| neutts-air | speech synthesis | Q8_0 | 748M | 1.90 GiB | 5.54 GiB |
| s2-proMoE | speech synthesis | Q8_0 | 4.6B | 6.07 GiB | 1.37 GiB |
| Fun-CosyVoice3-0.5B-2512 | speech synthesis | F16 | — | 4.37 GiB | 3.07 GiB |
| VoxCPM2 | speech synthesis | F16 | 2.3B | 5.57 GiB | 1.87 GiB |
| Kokoro-82M | speech synthesis | F16 | 82M | 1.00 GiB | 6.44 GiB |
| VieNeu-TTS | speech synthesis | Q4_0 | 553M | 1.54 GiB | 5.90 GiB |
| csm-1b | speech synthesis | F16 | 1.6B | 5.13 GiB | 2.31 GiB |
| VibeVoice-Realtime-0.5B | speech synthesis | F16 | 1.0B | 2.74 GiB | 4.70 GiB |
| MOSS-TTS-Local-Transformer-v1.5 | speech synthesis | Q8_0 | 4.6B | 6.41 GiB | 1.03 GiB |
| Qwen3-TTS-12Hz-1.7B-Base | speech synthesis | F16 | 1.9B | 4.44 GiB | 3.00 GiB |
| orpheus-3b-0.1-pretrained | speech synthesis | Q6_K_L | 3.8B | 7.43 GiB | 0.01 GiB |
| Qwen3-TTS-12Hz-1.7B-VoiceDesign | speech synthesis | F16 | 1.9B | 4.42 GiB | 3.02 GiB |
| VibeVoice-1.5B | speech synthesis | BF16 | 2.7B | 5.89 GiB | 1.55 GiB |
| VoxCPM-0.5B | speech synthesis | Q8_0 | 728M | 4.92 GiB | 2.52 GiB |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | speech synthesis | F16 | 1.9B | 4.42 GiB | 3.02 GiB |
This page models a generic 8GB 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.