Reson8 open-sources Rumble-2-Flash, a 0.5B speech model that matches a 30B model on 302 hours of call-centre audio
Synopsis
Reson8 has released Rumble-2-Flash under Apache 2.0, a 0.5B-parameter speech recognition model covering 40+ languages with automatic language detection. It was compressed from the larger production model Rumble-2 by on-policy distillation, in which the student transcribes audio itself and the teacher grades each token with its own probability. On the Open ASR Leaderboard multilingual track it averages a 6.96% error rate, better than Meta's 15x larger omniASR-LLM 7B; on AppTek's 302-hour, 14-accent call-centre benchmark it reaches 7.86% WER, on par with the 30B Nemotron 3.5 Transcribe at 7.78%.
Interpretation
A 0.5B-parameter speech recognition model is released under Apache 2.0, covering 40+ languages with automatic language detection, and the release states it competes with models 3 to 60 times its size. Comparable open models have been larger or needed custom runtime code; Rumble-2-Flash runs natively in Transformers and vLLM with no forked runtime or trust_remote_code. On the Open ASR Leaderboard multilingual track across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English it averages a 6.96% error rate, better than Meta's 15x larger omniASR-LLM 7B and ahead of Whisper large-v3; on the English track it averages 6.6% WER.
The model is compressed from the larger production model Rumble-2 by on-policy distillation: the student does not train on teacher-written transcripts but transcribes the audio itself, and the teacher supplies its own next-token probability at every position for the student to match. On-policy distillation normally requires a shared tokenizer, but choosing the Qwen3-ASR architecture means adopting its vocabulary, which differs from the teacher's; GOLD merges tokens the two tokenizers split differently so both models are compared over the same text, directly for shared tokens and by sorted probabilities for the rest. On AppTek's call-centre benchmark of 302 hours of helpline calls with 14 English accents, scored with AppTek's own script, Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA Nemotron 3.5 Transcribe (30B) at 7.78%, ahead of the same-family Qwen3-ASR 0.6B at 8.7%, and close to Qwen3-ASR 1.7B at 7.4%.
Perspective
The demonstrated scope is multilingual and English transcription on a public leaderboard plus AppTek's 302-hour, 14-accent call-centre benchmark; the throughput figure corresponds to the larger datasets on HF's Open ASR Leaderboard where batches stay full.
Throughput depends on how full the batches are, and the stated figure above 7,000 RTFx holds only when batches stay full; the call-centre benchmark is scored with AppTek's own script, and Qwen3-ASR 1.7B at 7.4% still does better in that comparison.
