Reson8 releases Rumble-2-Flash speech model under Apache 2.0
Reson8 released its 500-million-parameter Rumble-2-Flash speech-recognition model as open weights, with support for 45 languages.
Reson8 has released Rumble-2-Flash as an open-weights download, making the 500-million-parameter speech-recognition model available on Hugging Face under the Apache 2.0 license. Developers can run the model on their own infrastructure.
The repository identifies Rumble-2-Flash as a 0.5-billion-parameter BF16 model. The model card includes instructions for running it locally with Hugging Face Transformers or serving it through vLLM.
Reson8 lists support for 45 languages, including English, Arabic, Bengali, Cantonese, Chinese, French, German, Hindi, Japanese, Korean, Portuguese, Russian, Spanish, Thai, Turkish, Ukrainian and Vietnamese.
Reson8 says an Open ASR Leaderboard v1.16.1 run on an H200 using vLLM produced an average English word error rate of 6.60. In a separate company benchmark comparison, the company says Rumble-2-Flash averaged a 17.6% word error rate across 10 lower-resource languages, compared with 39.8% for Whisper Large V3. Reson8 reported those results, which were not independently reproduced for the research supporting this article.
Reson8’s model documentation lists Rumble-2-Flash for prerecorded transcription, but not for its Realtime or Turns endpoints. The company says the model produces plain transcripts without speaker labels, timestamps or vocabulary customization.
More news

Cohere previews timestamped video search in Compass

MBZUAI researchers constrain speech decoding to restore Arabic diacritics

Alibaba launches Qwen4-Max multimodal model across six regions
