All tasks / Voice & audio / Transcription
Best local model for transcription on a 24 GB GPU
Open-weights models whose approximate minimum GPU memory is 24 GB or less, ranked from public benchmarks. Memory figures are approximate; quantised builds can need less.
ATop ranked: Qwen3-ASR-1.7B (Alibaba (Qwen)), 5.4% on Hugging Face Open ASR Leaderboard (English, short-form average WER).14 options listed · data verified 19 Oct 2026
Ranked list
| Tier | Name | Approach | Runs | Rated by | Price | Skill |
|---|---|---|---|---|---|---|
| A |  Qwen3-ASR-1.7BAlibaba (Qwen) | Models | Local | 5.4% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~6 GB VRAM | — |
| A |  Higgs Audio v3 STTBoson AI | Models | Local | 5.9% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~6 GB VRAM | — |
| A |  Granite Speech 4.1 2BIBM | Models | Local | 6% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~6 GB VRAM | — |
| A |  MOSS-Transcribe-DiarizeOpenMOSS (Fudan University / MOSI) | Models | Local | 6.2% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~2 GB VRAM | — |
| A |  distil-large-v3.5Hugging Face | Models | Local | 6.8% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~2 GB VRAM | — |
| A |  Whisper large-v3OpenAI | Models | Local | 4.1% on Artificial Analysis Speech to Text: AA-WER v2 (non-streaming) | ~10 GB VRAM | — |
| B |  Canary-Qwen-2.5BNVIDIA | Models | Local | 4.3% on Artificial Analysis Speech to Text: AA-WER v2 (non-streaming) | ~6 GB VRAM | — |
| B |  Fun-ASR-Nano-2512Alibaba (FunAudioLLM) | Models | Local | 7.2% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~2 GB VRAM | — |
| B | Parakeet TDT 0.6B v3 NVIDIA | Models | Local | 4.5% on Artificial Analysis Speech to Text: AA-WER v2 (non-streaming) | ~2 GB VRAM | — |
| B |  cohere-transcribe-03-2026Cohere | Models | Local | 4.6% on Artificial Analysis Speech to Text: AA-WER v2 (non-streaming) | ~6 GB VRAM | — |
| B |  omniASR-LLM-7B-v2Meta (FAIR) | Models | Local | 8.1% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~16 GB VRAM | — |
| B |  VibeVoice-ASRMicrosoft | Models | Local | 8.2% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~19 GB VRAM | — |
| B |  Voxtral Mini 4B Realtime (2602)Mistral AI | Models | Local | 8.4% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~10 GB VRAM | — |
| C | Nemotron 3.5 ASR Streaming 0.6B NVIDIA | Models | Local | 9.8% on Hugging Face Open ASR Leaderboard (English, short-form average WER) | ~2 GB VRAM | — |
Tiers come only from public benchmarks (Hugging Face Open ASR Leaderboard (English, short-form average WER), Artificial Analysis Speech to Text: AA-WER v2 (non-streaming)); tools without a result of their own inherit the tier of the best model they offer. How we rank.