Everything you need to know before download
Small model, big ears
We wanted to see how small a model could be without the accuracy falling off. The target was a model small enough to run on everything, from voice notes to full call archives, while still holding up on complex, noisy audio.
The main challenge: real phone calls
Transcribing clean audio is easy, phone calls are not. The audio is squashed and noisy, with fast-talking people on opposite sides of the globe. So we ran rumble-2-flash on AppTek’s call-centre test, scored with AppTek’s own script.
It got about 92% of the words right. That’s on par with NVIDIA’s newest speech model, which is 60 times bigger, and ahead of Mistral, IBM, Cartesia, Deepgram, ElevenLabs and AssemblyAI. One model does better: Alibaba’s Qwen3-ASR 1.7B, more than three times our size.
Accuracy for its size
On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That's better than Meta's omniASR-LLM 7B, a model 15x larger, and ahead of Whisper large-v3. On the English track it averages 6.6% WER.
11 open-weight models, 9 languages (1 above 12% not shown). Average error rate against model size. Higher up is lower error.
Average over Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English.
Hugging Face Open ASR Leaderboard, 13 November 2026. Chinese is scored by character error rate.
The best score for its size
Speed
Rumble-2-Flash is also fast. Its throughput depends on how full the batches are: on the larger datasets in HF's Open ASR Leaderboard, where batches stay full, it reaches over 7,000 RTFx. That's an hour of audio in about half a second.
Languages around the globe
Large models are mostly trained on English and a handful of major languages, and their error rates climb quickly outside that group.
We compared Rumble-2-Flash with OpenAI’s Whisper Large V3, a model three times its size, on ten of these languages. Rumble-2-Flash averages 17.6% WER, compared with 39.8% for Whisper, which is less than half the errors. The gap is widest where other models struggle most: 13% vs. 72% in Maltese, 26% vs. 74% in Bangla, and 13% vs. 42% in Basque.
What you need to know about benchmarking
Speech tests compare what a model writes with what a person typed out, and people make mistakes too. They mishear, or write “twenty” where the model wrote “20”, and the test counts that as wrong.
That’s how anyone’s scores should be read, ours included. If you’re evaluating speech-to-text for your own product, test it on your own audio.
The model is yours.
Are you ready 2 Rumble?
Run it
Laptop, on-prem server, device or your own cloud. Open weights so audio never has to leave your hardware.
Own it
Licensed under Apache 2.0. Use it commercially, build it into your product.
Shape it
Fine-tune for your needs, accents, terminology or language. Small enough to train without a cluster.
For the rooms where every word matters
Even in a busy meeting or a noisy room, Rumble-2-Flash open weights gets around 9 in 10 words right. That’s plenty for a voice note. In a doctor’s appointment, a board meeting or a voice agent serving millions of callers, one wrong word in that last 1 in 10 can change a diagnosis or a decision. That’s what Rumble-2 is for: the most accurate model we’ve trained so far.
