Reson8 has open-sourced Rumble-2-Flash under the Apache 2.0 licence: a 0.5-billion-parameter speech recognition model covering more than 40 languages that its makers say competes with models three to sixty times its size.
The model runs natively in Hugging Face Transformers and vLLM with no custom code, detects languages automatically, and gives control over output formatting, including numbers as digits or words, fillers, verbatim repeats and forced language.
On accuracy, Rumble-2-Flash averages a 6.96 percent error rate on the multilingual track of Hugging Face's Open ASR Leaderboard, covering Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That puts it ahead of Meta's omniASR-LLM 7B, a model fifteen times larger, and ahead of Whisper large-v3. On the English track it averages 6.6 percent word error rate.
It is also fast. On the larger leaderboard datasets, where batches stay full, it reaches more than 7,000 RTFx — an hour of audio transcribed in about half a second.
It holds up on real phone calls, too. On AppTek's call-centre benchmark of 302 hours of helpline calls across 14 English accents, it reaches 7.86 percent word error rate, on par with NVIDIA's Nemotron 3.5 Transcribe, a 30-billion-parameter model, at 7.78 percent.
Rumble-2-Flash was built by distilling Reson8's larger production model, Rumble-2, using on-policy distillation on the Qwen3-ASR architecture. A technique the team calls GOLD aligns the two models' different tokenizers so the student can learn from the teacher token by token.
The weights are available on the Hugging Face Hub as Reson8/Rumble-2-Flash, and through Reson8's API.




