Reson8 Releases Rumble-2-Flash: Open 0.5B Speech Model for 45 Languages
Reson8 launches Rumble-2-Flash, a 0.5B parameter multilingual speech recognition model supporting 45 languages with controllable output formatting, available under Apache 2.0 on Hugging Face.

By Clara Tan, Artificial Intelligence
, updated 3 min read
Reson8 has released Rumble-2-Flash, a 0.5 billion parameter multilingual automatic speech recognition model, on Hugging Face under an Apache 2.0 licence. The model supports 45 languages and introduces controllable output formatting through special tokens.
What the model does
Rumble-2-Flash is built on the Qwen3-ASR architecture and uses a Resonant-Light chat template that accepts control tokens for output formatting. Users can specify whether numbers appear as digits or words, whether to keep or drop filler words like "um" and "uh", whether to preserve or collapse repeated words, and can force transcription into a specific language.
The model card lists support for 45 languages including English, Chinese, Spanish, French, German, Japanese, Korean, Arabic, Hindi, Portuguese, Russian, Italian, Dutch, Polish, Turkish, Vietnamese, Thai, Indonesian, Swedish, Danish, Norwegian, Finnish, Greek, Czech, Hungarian, Romanian, Hebrew, Ukrainian, Catalan, Croatian, Serbian, Slovak, Slovenian, Estonian, Latvian, Lithuanian, Bulgarian, Macedonian, Maltese, Icelandic, Welsh, Basque, Galician and Luxembourgish.
How to use it
The model integrates with Hugging Face Transformers and vLLM. A minimal example using the Transformers pipeline:
from transformers import pipeline
pipe = pipeline("automatic-speech-recognition", model="Reson8/Rumble-2-Flash")
result = pipe("audio.wav")For controllable output, pass control tokens in the system message or use the chat template with parameters for digits, filler, verbatim and language. A Gradio demo is available on the model page for quick testing without code.
Benchmark results
Reson8 reports word error rate (WER) results on four public benchmarks:
Common Voice 15: 4.2% WER averaged across 25 languages
MLS (Multilingual LibriSpeech): 3.8% WER on 8 languages
VoxPopuli: 5.1% WER on 15 languages
Fleurs: 6.3% WER on 30 languages
Full per-language breakdowns are in the model card on Hugging Face.
Why it matters
Small open-weight speech models that run locally on consumer hardware are valuable for privacy-sensitive applications, offline transcription, and developers who need controllable output formatting without API costs. The 0.5B parameter size means Rumble-2-Flash can run on a modern CPU or modest GPU, and the Apache 2.0 licence permits commercial use.
The controllable output tokens address a practical gap: most ASR models produce raw transcripts that need post-processing for formatting. Building digits, filler handling and verbatim control into the model itself simplifies pipelines for captioning, meeting notes and voice interfaces.
Source
Hugging Face blog announcement: https://huggingface.co/blog/rumble-2-flash
Model card: https://huggingface.co/Reson8/Rumble-2-Flash
Sources
- 1Hugging Face blog announcementhuggingface.co
- 2Model card on Hugging Facehuggingface.co