{"slug": "openai-s-whisper-large-v3-runs-locally-under-1gb-without-python", "title": "OpenAI's Whisper large-v3 Runs Locally Under 1GB Without Python", "summary": "A GGUF-packaged build of OpenAI's 1.5-billion-parameter Whisper large-v3, published by handy-computer for the new ggml-based transcribe.cpp runtime, reaches 1.81% word error rate on LibriSpeech test-clean at F16, beating OpenAI's self-reported 2.01%, and its smallest Q4_K_M file runs locally in 997 MB. The repository offers five quantizations from F16 (3.11 GB) to Q4_K_M (997 MB) with only a 0.05-percentage-point WER spread, runs on Metal, Vulkan, CUDA and CPU, and had recorded more than 829,000 downloads and 3,100 likes on Hugging Face under an Apache 2.0 license. The conversion keeps transcription across roughly 100 languages, language identification, segment timestamps and translation into English, giving developers a compact multilingual transcription option without a hosted API or Python runtime.", "body_md": "# OpenAI's Whisper large-v3 Runs Locally Under 1GB Without Python\n\nA GGUF-packaged Whisper large-v3 built for transcribe.cpp hits 1.81% WER on LibriSpeech and runs on a laptop in under 1 GB.\n\n- Whisper large-v3 repackaged as GGUF for the new [transcribe.cpp](https://github.com/handy-computer/transcribe.cpp?utm_source=alphasignal) runtime, 829K+ downloads\n- Hits 1.81% WER on LibriSpeech test-clean at F16, beating OpenAI's self-reported 2.01%\n- Five quants from F16 (3.11 GB) down to Q4_K_M (997 MB), WER spread only 0.05 points\n- Runs on Metal, Vulkan, CUDA, and CPU via ggml with memory-mapped weights for fast loading\n- Apache-2.0 license, supports 100 languages plus any-to-English translation and timestamps\n- Turbo variant at 886 MB available via [handy-computer](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf?utm_source=alphasignal) for speed-first use cases\n\nThe [model repository](https://huggingface.co/handy-computer/whisper-large-v3-gguf?utm_source=alphasignal) packages OpenAI’s 1.5-billion-parameter [Whisper large-v3](https://huggingface.co/openai/whisper-large-v3?utm_source=alphasignal) as GGUF files for [transcribe.cpp](https://github.com/handy-computer/transcribe.cpp?utm_source=alphasignal), a new ggml-based inference engine. The smallest file occupies 997 MB and runs locally across several hardware backends, giving developers a compact multilingual transcription option without a hosted API or Python runtime.\n\nGGUF stores model tensors and metadata in a single file designed for efficient loading by ggml runtimes. Quantization reduces the precision of those tensors, shrinking storage and memory requirements at the cost of some numerical fidelity. The repository reports only a 0.05-percentage-point increase in word error rate between its F16 and Q4_K_M files.\n\nAt publication time, the repository had recorded more than 829,000 downloads and 3,100 likes on Hugging Face. Its model page lists the files under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0?utm_source=alphasignal), which permits commercial use subject to the license terms and notice requirements.\n\n## One model, five memory budgets\n\nThe conversion preserves large-v3’s core capabilities: transcription across roughly 100 languages, automatic language identification, segment timestamps, and speech translation into English. Compared with earlier Whisper checkpoints, large-v3 processes 128-bin log-Mel spectrograms instead of 80-bin inputs and includes a Cantonese language token.\n\nThe reference file stores weights in 16-bit floating point. Four quantized versions reduce the file size by representing most weights with fewer bits, with the K-series formats applying mixed precision to protect selected tensors.\n\n| Published file sizes and LibriSpeech test-clean results |  |  | \n|---|---|---|\n| Quantization | File size | Word error rate | \n|---|---|---|\n| F16 | 3.11 GB | 1.81% | \n| Q8_0 | 1.67 GB | 1.82% | \n| Q6_K | 1.30 GB | 1.83% | \n| Q5_K_M | 1.16 GB | 1.84% | \n| Q4_K_M | 997 MB | 1.86% | \n\nWord error rate counts substitutions, deletions, and insertions against a reference transcript; lower values indicate fewer errors. LibriSpeech test-clean consists of clean English audiobook speech, so these results provide little evidence about noisy recordings, far-field microphones, accents, or multilingual accuracy.\n\n## This story is for Pro members\n\nYou've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.", "url": "https://wpnews.pro/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python", "canonical_source": "https://alphasignal.ai/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python", "published_at": "2026-11-15 16:00:31+00:00", "updated_at": "2026-11-15 16:23:05.383815+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-tools", "ai-infrastructure", "natural-language-processing"], "entities": ["OpenAI", "Whisper large-v3", "transcribe.cpp", "handy-computer", "Hugging Face", "GGUF", "ggml", "LibriSpeech"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python", "markdown": "https://wpnews.pro/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python.md", "text": "https://wpnews.pro/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python.txt", "jsonld": "https://wpnews.pro/news/openai-s-whisper-large-v3-runs-locally-under-1gb-without-python.jsonld"}}