Reson8 has openly released its Rumble-2-Flash speech recognition model under the Apache 2.0 license. The 500-million-parameter system supports more than 40 languages and automatically detects the language being spoken. The company says it is competitive with substantially larger models.
Users can control, among other things, whether numbers appear as digits or words and whether filler words and verbatim repetitions are included in transcripts. A specific language can also be selected instead of using automatic detection.
A smaller model trained by a larger one
Rumble-2-Flash is built on the Qwen3-ASR architecture and has native support in the Transformers and vLLM frameworks. Reson8 says it requires neither custom code nor a modified runtime.
The company transferred capabilities from its larger production model, Rumble-2, to the smaller system using a method called on-policy distillation. The smaller model produces its own transcript of the audio, while the teacher model provides token-by-token feedback on the continuation it considers probable. The student therefore receives feedback on its own outputs and mistakes.
The process was complicated by the models’ different tokenizers. The GOLD method makes their outputs comparable by merging tokens where the text has been split differently. Tokens present in both vocabularies are compared directly, while the others are compared using their sorted probabilities.
Benchmark results and speed
According to Reson8, Rumble-2-Flash averaged a 6.96% error rate on the multilingual track of the Open ASR Leaderboard across nine languages: Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The company says that result is better than those of Meta’s 7-billion-parameter omniASR-LLM model and Whisper large-v3. On the English track, the model averaged a 6.6% word error rate.
Processing speed depends on how full the batches are. The company says the model exceeds 7,000 RTFx on larger benchmark datasets with full batches, corresponding to roughly half a second to process an hour of audio.
Tested on phone calls
Compressed, noisy phone calls and differing accents pose particular challenges for speech recognition. On AppTek’s call-centre benchmark, which contains 302 hours of helpline calls across 14 English accents, Rumble-2-Flash achieved a 7.86% word error rate using AppTek’s own scoring script.
That was close to the 7.78% result recorded by NVIDIA’s 30-billion-parameter Nemotron 3.5 Transcribe. In the same test, Qwen3-ASR 0.6B scored 8.7% and Qwen3-ASR 1.7B scored 7.4%; the latter was the only model in the comparison to outperform Rumble-2-Flash.
The model weights are available under the identifier Reson8/Rumble-2-Flash, and the system can also be used through the company’s API.



