Reson8 Open-Sources Rumble-2-Flash, 0.5B Speech Model It Says Rivals 30B Nemotron
Summary
Reson8 open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model covering more than 40 languages, under the Apache 2.0 license. The company reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track and 7.86% word error rate on AppTek's 302-hour call-centre benchmark, close to NVIDIA's 30B Nemotron 3.5 Transcribe.
Key Points
- Reson8 reports over 7,000 RTFx on full batches, roughly an hour of audio in half a second.
- Reson8 builds Rumble-2-Flash by distilling its larger Rumble-2 model with on-policy distillation plus GOLD, which merges tokens across differing tokenizers.
- Reson8 publishes the weights on Hugging Face at Reson8/Rumble-2-Flash and also serves the model through its own API.