Source: huggingface.co/blog/rumble-2-flash
Rumble-2-Flash detects language automatically when none is given and can output numbers as digits or words, keep or drop fillers, keep verbatim repeats, and force a language. Reson8 says it transcribes about an hour of audio in roughly half a second, with throughput above 7,000 RTFx on full batches in Hugging Face's Open ASR Leaderboard.
The model is built on the Qwen3-ASR architecture and was distilled from Reson8's larger production model, Rumble-2, using on-policy distillation. Reson8 used its GOLD method so that teacher and student could have different tokenizers, with the teacher scoring the student's own transcripts token by token.
On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across nine languages and 6.6% WER on the English track. On AppTek's 302-hour call-centre benchmark with 14 English accents, it reaches 7.86% WER, close to NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%.
Why it matters if you build agents
Developers can run speech transcription in an existing Transformers or vLLM stack without custom code or trust_remote_code, which makes a small self-hosted option practical for batch jobs.
The call-centre results cover compressed, accented phone audio, a harder case than clean benchmarks, so the figures are a useful starting point. Agents handling phone calls should still test the model on their own recordings before relying on it.
Get stories like this every weekday
One short email with the day's agent tooling news.