From the source
- Reson8 has open sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model for 40+ languages, under the Apache 2.0 license.
- The model is distilled from Rumble-2, the company's larger production model, using on-policy distillation, where the student transcribes audio itself and the teacher grades each token rather than supplying transcripts.
- Because the two models use different tokenizers, the team used GOLD to merge tokens where the tokenizers split text differently so both are compared over the same text.
- On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across nine languages, which the company says beats
Meta's omniASR-LLM 7B, a model 15x larger, and Whisper large-v3.
- On AppTek's 302-hour call-centre benchmark with 14 English accents, it reaches 7.86% WER, on par with
NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%.
- The weights are on the Hub at Reson8/Rumble-2-Flash and it is also available through the company's API.





