Reson8 has released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model, under the Apache 2.0 licence, which allows commercial use and modification. The company says the model transcribes speech in more than 40 languages, detects the language automatically, and runs through Hugging Face Transformers or vLLM without custom code. Users can also control output formatting, including whether numbers appear as digits or words, whether filler words are kept, and whether verbatim repeats are retained.
The model was built by distilling Reson8's larger production model, Rumble-2, into a smaller student using on-policy distillation. Instead of learning from fixed transcripts, the student produces its own transcripts and receives feedback on each token from the teacher. Because the two models use different tokenisers, the team used a method called GOLD to compare them over the same text. On the Open ASR Leaderboard's multilingual track, Rumble-2-Flash averages a 6.96% word error rate across nine languages, which the company says beats Meta's 7-billion-parameter omniASR-LLM and Whisper large-v3. On a 302-hour AppTek call-centre benchmark with 14 English accents, it scored 7.86%, close to NVIDIA's 30-billion-parameter Nemotron 3.5 Transcribe at 7.78%, and behind only Qwen3-ASR 1.7B at 7.4%. The company reports throughput above 7,000 RTFx on full batches, equivalent to about half a second to transcribe an hour of audio. These results are self-reported, and the call-centre test used AppTek's own scoring script.
For Australian businesses and professionals, the release changes little immediately unless they already transcribe calls, meetings or customer recordings. The weights are free, but running them requires suitable GPU hardware or cloud compute. The benchmarks do not say whether Australian English was included, so accuracy on local accents and phone audio should be tested before relying on it. If recordings contain personal information, the Privacy Act 1988 and the Australian Privacy Principles apply to how they are collected, stored and disclosed, whichever model is used. Self-hosting may simplify data handling compared with sending audio to a third-party API, but it does not remove those obligations. A small trial on your own recordings is worth the effort before changing any workflow.
Key Points
Rumble-2-Flash is a 0.5B-parameter open-weight speech recognition model released under Apache 2.0, supporting 40+ languages with automatic language detection.
It was distilled from the larger Rumble-2 using on-policy distillation, with a GOLD method to handle mismatched tokenisers.
Reson8 reports a 6.96% average word error rate on the Open ASR multilingual track and 7.86% on a 302-hour call-centre benchmark, close to a 30B model.
The model runs natively in Hugging Face Transformers and vLLM and reports over 7,000 RTFx on full batches.
The benchmark results are vendor-reported, and the article does not confirm whether Australian English was tested.