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. 29 indexed models fit at 32K context — the largest being canary-qwen-2.5b at 2.6B parameters in Q6_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◐ |
|---|---|---|---|---|---|
| nemotron-3.5-asr-streaming-0.6b | speech recognition | F32 | 638M | 3.22 GiB | 0.50 GiB |
| parakeet-tdt-0.6b-v3 | speech recognition | F32 | 627M | 3.18 GiB | 0.54 GiB |
| whisper-medium | speech recognition | F32 | 764M | 3.69 GiB | 0.03 GiB |
| whisper-large-v3 | speech recognition | Q8_0 | 1.5B | 2.44 GiB | 1.28 GiB |
| whisper-large-v3-turbo | speech recognition | F16 | 809M | 2.36 GiB | 1.36 GiB |
| Qwen3-ASR-1.7B | speech recognition | Q8_0 | 2.3B | 3.18 GiB | 0.54 GiB |
| Qwen3-ASR-0.6B | speech recognition | F16 | 938M | 2.32 GiB | 1.40 GiB |
| GigaAM-v3 | speech recognition | F32 | 223M | 1.67 GiB | 2.05 GiB |
| parakeet-ctc-0.6b | speech recognition | F32 | 609M | 3.11 GiB | 0.61 GiB |
| whisper-small | speech recognition | F32 | 242M | 1.75 GiB | 1.97 GiB |
| whisper-large | speech recognition | Q8_0 | 1.5B | 2.40 GiB | 1.32 GiB |
| canary-1b-flash | speech recognition | F16 | 811M | 2.51 GiB | 1.21 GiB |
| whisper-large-v2 | speech recognition | Q8_0 | 1.5B | 2.40 GiB | 1.32 GiB |
| canary-qwen-2.5b | speech recognition | Q6_K | 2.6B | 2.90 GiB | 0.82 GiB |
| Breeze-ASR-25 | speech recognition | Q8_0 | 1.5B | 2.40 GiB | 1.32 GiB |
| nemotron-speech-streaming-en-0.6b | speech recognition | F32 | 618M | 3.15 GiB | 0.57 GiB |
| parakeet-ctc-1.1b | speech recognition | F16 | 1.1B | 2.83 GiB | 0.89 GiB |
| parakeet-rnnt-1.1b | speech recognition | F16 | 1.1B | 2.84 GiB | 0.88 GiB |
| whisper-base | speech recognition | F32 | 73M | 1.12 GiB | 2.60 GiB |
| moonshine-streaming-medium | speech recognition | F32 | 266M | 2.85 GiB | 0.87 GiB |
| whisper-medium.en | speech recognition | F32 | 764M | 3.69 GiB | 0.03 GiB |
| parakeet-rnnt-0.6b | speech recognition | F32 | 617M | 3.14 GiB | 0.58 GiB |
| whisper-small.en | speech recognition | F32 | 242M | 1.75 GiB | 1.97 GiB |
| moonshine-streaming-small | speech recognition | F32 | 140M | 1.91 GiB | 1.81 GiB |
| whisper-tiny | speech recognition | F32 | 38M | 0.99 GiB | 2.73 GiB |
| whisper-base.en | speech recognition | F32 | 73M | 1.12 GiB | 2.60 GiB |
| moonshine-streaming-tiny | speech recognition | F32 | 44M | 1.16 GiB | 2.56 GiB |
| moonshine-base | speech recognition | F32 | 62M | 0.99 GiB | 2.73 GiB |
| Qwen3-ForcedAligner-0.6B | speech recognition | F16 | 918M | 2.56 GiB | 1.16 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.