Reson8 announced on November 18, 2026 that it has open sourced Rumble-2-Flash under the Apache 2.0 license. The model contains 0.5B parameters, supports automatic language detection for more than 40 languages, and offers output controls such as rendering numbers as digits or words, preserving fillers and verbatim repeats, and forcing a target language.
Architecture and training
Rumble-2-Flash uses the Qwen3-ASR architecture and is distributed for immediate use in Transformers and vLLM without requiring custom runtime code. Reson8 says the model was produced by distilling their larger production model, Rumble-2, into a smaller student via on-policy distillation. In this process the student transcribes audio itself while the teacher assigns token-by-token probabilities to correct the student’s outputs.
To address tokenizer differences between teacher and student, Reson8 applied a method called GOLD. GOLD merges tokens where tokenizers split text differently so that comparisons are made over a common text: shared tokens are compared directly, and others by sorted probabilities.
Benchmarks and speed
On the Open ASR Leaderboard multilingual track, Reson8 reports that Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. Reson8 states this result is better than Meta’s omniASR-LLM 7B (15x larger) and ahead of Whisper large-v3; on the English track Rumble-2-Flash averages 6.6% WER.
Reson8 also highlights runtime performance: on large datasets with full batches the model achieves over 7,000 RTFx, which the team equates to transcribing an hour of audio in about half a second. On AppTek’s 302-hour call-centre benchmark (14 English accents), Rumble-2-Flash reached 7.86% WER, compared with NVIDIA’s Nemotron 3.5 Transcribe (30B) at 7.78%. Reson8 notes Qwen3-ASR 0.6B scored 8.7% and Qwen3-ASR 1.7B scored 7.4% in the same comparison.
Weights are available on the Hub at Reson8/Rumble-2-Flash and the model is offered through Reson8’s API.
Original source: Hugging Face Blog