Rumble-2-Flash: Open 0.5B Speech Model Transcribes 40+ Languages at 7,000 RTFx

2 min read
Rumble-2-Flash: Open 0.5B Speech Model Transcribes 40+ Languages at 7,000 RTFx
Reson8 releases Rumble-2-Flash under Apache 2.0 — a tiny 0.5B-parameter speech recognition model that matches 30B-parameter models on real phone calls and runs anywhere with native Transformers and vLLM support.

A Tiny Model That Punches Far Above Its Weight

Reson8 has open-sourced Rumble-2-Flash, a 0.5-billion-parameter automatic speech recognition model supporting over 40 languages with automatic language detection. Despite its small size, it achieves a 6.96% average error rate on the Open ASR Leaderboard multilingual track — better than Meta's 7B-parameter omniASR-LLM and ahead of Whisper large-v3. On English, it hits 6.6% WER.

Built on Qwen3-ASR With On-Policy Distillation

Rumble-2-Flash uses the Qwen3-ASR architecture, giving it day-zero support in Hugging Face Transformers and vLLM with no custom code or trust_remote_code required. To shrink the model without losing accuracy, Reson8 distilled its larger production model (Rumble-2) using on-policy distillation: the student model transcribes audio itself, and the teacher grades every token position. A technique called GOLD aligns the two models' different tokenizers so they can be compared token-by-token.

Real-World Performance on Noisy Phone Calls

Clean audio is easy; phone calls are not. On AppTek's 302-hour call-center benchmark with 14 English accents, Rumble-2-Flash reaches 7.86% WER — on par with NVIDIA's 30-billion-parameter Nemotron 3.5 Transcribe at 7.78%. It outperforms Qwen3-ASR 0.6B (8.7%) and nearly matches Qwen3-ASR 1.7B (7.4%), the only smaller model that beats it.

Speed: An Hour of Audio in Half a Second

Throughput scales with batch fullness. On the larger Open ASR Leaderboard datasets where batches stay full, Rumble-2-Flash exceeds 7,000 RTFx — transcribing an hour of audio in about half a second. The model runs on CPU or GPU and integrates directly into existing Transformers or vLLM pipelines.

Output Control and Flexibility

Rumble-2-Flash offers practical output controls: numbers as digits or words, filler word handling, verbatim repeat options, and forced language selection. Audio input accepts local paths, URLs, NumPy arrays, or lists of these, making it flexible for batch processing and streaming use cases.

Available Now on Hugging Face

The weights are on the Hugging Face Hub at Reson8/Rumble-2-Flash under the Apache 2.0 license. An API is also available. Reson8 invites developers to share what they build in the Discussions tab.

Found this useful?

Share it with someone shopping for a phone.

Keep reading

All articles