Best local AI models for 96GB VRAM
Ranked by what actually fits at 32K context, computed from real file bytes.
A 96GB card gives you about 89.28 GiB to work with after driver overhead. 39 indexed models fit at 32K context — the largest being Voxtral-Small-24B-2507 at 24.3B parameters in F16.
From the file· fit from summed bytesFrom the file· KV per layer
Fits in 96GB 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 | 86.06 GiB |
| parakeet-tdt-0.6b-v3 | speech recognition | F32 | 627M | 3.18 GiB | 86.10 GiB |
| whisper-medium | speech recognition | F32 | 764M | 3.69 GiB | 85.59 GiB |
| Voxtral-Mini-4B-Realtime-2602 | speech recognition | F16 | 4.4B | 12.33 GiB | 76.95 GiB |
| whisper-large-v3 | speech recognition | F16 | 1.5B | 3.74 GiB | 85.54 GiB |
| whisper-large-v3-turbo | speech recognition | F16 | 809M | 2.36 GiB | 86.92 GiB |
| Qwen3-ASR-1.7B | speech recognition | F16 | 2.3B | 5.23 GiB | 84.05 GiB |
| Voxtral-Small-24B-2507 | speech recognition | F16 | 24.3B | 51.13 GiB | 38.15 GiB |
| Qwen3-ASR-0.6B | speech recognition | F16 | 938M | 2.32 GiB | 86.96 GiB |
| GigaAM-v3 | speech recognition | F32 | 223M | 1.67 GiB | 87.61 GiB |
| parakeet-ctc-0.6b | speech recognition | F16 | 609M | 16.97 GiB | 72.31 GiB |
| granite-speech-4.1-2b-nar | speech recognition | F16 | 2.3B | 8.65 GiB | 80.63 GiB |
| whisper-small | speech recognition | F32 | 242M | 1.75 GiB | 87.53 GiB |
| granite-speech-4.1-2b | speech recognition | F16 | 2.3B | 8.48 GiB | 80.80 GiB |
| whisper-large | speech recognition | F32 | 1.5B | 6.60 GiB | 82.68 GiB |
| Voxtral-Mini-3B-2507 | speech recognition | F16 | 4.7B | 13.29 GiB | 75.99 GiB |
| canary-1b-flash | speech recognition | F32 | 811M | 4.16 GiB | 85.12 GiB |
| whisper-large-v2 | speech recognition | F32 | 1.5B | 6.60 GiB | 82.68 GiB |
| canary-qwen-2.5b | speech recognition | F16 | 2.6B | 6.15 GiB | 83.13 GiB |
| granite-speech-4.1-2b-plus | speech recognition | F16 | 2.1B | 8.50 GiB | 80.78 GiB |
| Breeze-ASR-25 | speech recognition | F16 | 1.5B | 3.74 GiB | 85.54 GiB |
| nemotron-speech-streaming-en-0.6b | speech recognition | F32 | 618M | 3.15 GiB | 86.13 GiB |
| granite-4.0-1b-speech | speech recognition | F16 | 2.3B | 7.60 GiB | 81.68 GiB |
| parakeet-ctc-1.1b | speech recognition | F32 | 1.1B | 4.80 GiB | 84.48 GiB |
| parakeet-rnnt-1.1b | speech recognition | F32 | 1.1B | 4.83 GiB | 84.45 GiB |
| whisper-base | speech recognition | F32 | 73M | 1.12 GiB | 88.16 GiB |
| moonshine-streaming-medium | speech recognition | F32 | 266M | 2.85 GiB | 86.43 GiB |
| whisper-medium.en | speech recognition | F32 | 764M | 3.69 GiB | 85.59 GiB |
| parakeet-rnnt-0.6b | speech recognition | F32 | 617M | 3.14 GiB | 86.14 GiB |
| whisper-small.en | speech recognition | F32 | 242M | 1.75 GiB | 87.53 GiB |
| moonshine-streaming-small | speech recognition | F32 | 140M | 1.91 GiB | 87.37 GiB |
| whisper-tiny | speech recognition | F32 | 38M | 0.99 GiB | 88.29 GiB |
| MOSS-Transcribe-Diarize | speech recognition | F32 | 909M | 7.66 GiB | 81.62 GiB |
| GLM-ASR-Nano-2512 | speech recognition | BF16 | 2.3B | 5.51 GiB | 83.77 GiB |
| ARK-ASR-3B | speech recognition | F16 | 4.1B | 8.93 GiB | 80.35 GiB |
| whisper-base.en | speech recognition | F32 | 73M | 1.12 GiB | 88.16 GiB |
| moonshine-streaming-tiny | speech recognition | F32 | 44M | 1.16 GiB | 88.12 GiB |
| moonshine-base | speech recognition | F32 | 62M | 0.99 GiB | 88.29 GiB |
| Qwen3-ForcedAligner-0.6B | speech recognition | F16 | 918M | 2.56 GiB | 86.72 GiB |
This page models a generic 96GB 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.