Best local AI models for 48GB VRAM

Ranked by what actually fits at 32K context, computed from real file bytes.

A 48GB card gives you about 44.64 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 Q8_0.

From the file· fit from summed bytesFrom the file· KV per layer

Fits in 48GB at 32K context

largest quantization that fits, per model
ModelModalityBest quantParamsTotalHeadroom
nemotron-3.5-asr-streaming-0.6bspeech recognitionF32638M3.22 GiB41.42 GiB
parakeet-tdt-0.6b-v3speech recognitionF32627M3.18 GiB41.46 GiB
whisper-mediumspeech recognitionF32764M3.69 GiB40.95 GiB
Voxtral-Mini-4B-Realtime-2602speech recognitionF164.4B12.33 GiB32.31 GiB
whisper-large-v3speech recognitionF161.5B3.74 GiB40.90 GiB
whisper-large-v3-turbospeech recognitionF16809M2.36 GiB42.28 GiB
Qwen3-ASR-1.7Bspeech recognitionF162.3B5.23 GiB39.41 GiB
Voxtral-Small-24B-2507speech recognitionQ8_024.3B29.96 GiB14.68 GiB
Qwen3-ASR-0.6Bspeech recognitionF16938M2.32 GiB42.32 GiB
GigaAM-v3speech recognitionF32223M1.67 GiB42.97 GiB
parakeet-ctc-0.6bspeech recognitionF16609M16.97 GiB27.67 GiB
granite-speech-4.1-2b-narspeech recognitionF162.3B8.65 GiB35.99 GiB
whisper-smallspeech recognitionF32242M1.75 GiB42.89 GiB
granite-speech-4.1-2bspeech recognitionF162.3B8.48 GiB36.16 GiB
whisper-largespeech recognitionF321.5B6.60 GiB38.04 GiB
Voxtral-Mini-3B-2507speech recognitionF164.7B13.29 GiB31.35 GiB
canary-1b-flashspeech recognitionF32811M4.16 GiB40.48 GiB
whisper-large-v2speech recognitionF321.5B6.60 GiB38.04 GiB
canary-qwen-2.5bspeech recognitionF162.6B6.15 GiB38.49 GiB
granite-speech-4.1-2b-plusspeech recognitionF162.1B8.50 GiB36.14 GiB
Breeze-ASR-25speech recognitionF161.5B3.74 GiB40.90 GiB
nemotron-speech-streaming-en-0.6bspeech recognitionF32618M3.15 GiB41.49 GiB
granite-4.0-1b-speechspeech recognitionF162.3B7.60 GiB37.04 GiB
parakeet-ctc-1.1bspeech recognitionF321.1B4.80 GiB39.84 GiB
parakeet-rnnt-1.1bspeech recognitionF321.1B4.83 GiB39.81 GiB
whisper-basespeech recognitionF3273M1.12 GiB43.52 GiB
moonshine-streaming-mediumspeech recognitionF32266M2.85 GiB41.79 GiB
whisper-medium.enspeech recognitionF32764M3.69 GiB40.95 GiB
parakeet-rnnt-0.6bspeech recognitionF32617M3.14 GiB41.50 GiB
whisper-small.enspeech recognitionF32242M1.75 GiB42.89 GiB
moonshine-streaming-smallspeech recognitionF32140M1.91 GiB42.73 GiB
whisper-tinyspeech recognitionF3238M0.99 GiB43.65 GiB
MOSS-Transcribe-Diarizespeech recognitionF32909M7.66 GiB36.98 GiB
GLM-ASR-Nano-2512speech recognitionBF162.3B5.51 GiB39.13 GiB
ARK-ASR-3Bspeech recognitionF164.1B8.93 GiB35.71 GiB
whisper-base.enspeech recognitionF3273M1.12 GiB43.52 GiB
moonshine-streaming-tinyspeech recognitionF3244M1.16 GiB43.48 GiB
moonshine-basespeech recognitionF3262M0.99 GiB43.65 GiB
Qwen3-ForcedAligner-0.6Bspeech recognitionF16918M2.56 GiB42.08 GiB
Spec sheetPredictedwhat these mean

This page models a generic 48GB 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.