Skip to content
Neural Daily

ModelsNews

Rumble-2-Flash: open 0.5B speech model for 40+ languages

🇭🇺 Olvasd el magyarul →

Reson8 released Rumble-2-Flash under Apache 2.0. The 0.5B ASR model handles 40+ languages, matches the 30B Nemotron 3.5 Transcribe on phone calls and transcribes an hour of audio in about half a second.

· 2 min read

Rumble-2-Flash announcement graphic: two smiling emoji faces with a sound wave between them on a purple background
Image: Hugging Face · Press kit · Official blog thumbnail from huggingface.co · Source

What happened?

On November 18 Reson8 open-sourced Rumble-2-Flash under the Apache 2.0 license (Hugging Face blog). It is a 0.5B-parameter speech recognition (ASR) model covering 40+ languages with automatic language detection. Weights are on the Hub (Reson8/Rumble-2-Flash) and the model is also available through the company's API.

What it can do

  • Open ASR Leaderboard, multilingual track: 6.96% average word error rate (WER) across nine languages (Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, English). That beats Meta's omniASR-LLM 7B, a model 15x larger, and Whisper large-v3.
  • English track: 6.6% average WER.
  • Speed: over 7,000 RTFx on full batches, about half a second for an hour of audio.
  • Real phone calls (AppTek call-centre benchmark, 302 hours, 14 English accents): 7.86% WER, versus 7.78% for NVIDIA's 30B Nemotron 3.5 Transcribe, 8.7% for Qwen3-ASR 0.6B and 7.4% for Qwen3-ASR 1.7B.
  • Output control: numbers as digits or words, fillers, verbatim repeats, forced language.

How it was built

The model uses the Qwen3-ASR architecture, so it runs natively in Transformers and vLLM with no custom code or trust_remote_code. It was distilled on-policy from Reson8's larger production Rumble-2 model: the student transcribes the audio itself and the teacher grades that transcript token by token, so the student is corrected on exactly the mistakes it makes. Because the two models use different tokenizers, the team used GOLD, which aligns tokens that the tokenizers split differently.

Why it matters

A half-billion-parameter model is cheap to run on your own hardware, which matters for on-prem transcription, call-centre analytics or audio that has to stay in-house. On the phone-call test it matches the 60x larger Nemotron 3.5 Transcribe. The numbers are the company's own, so test it on your own audio before replacing an existing ASR stack. For independent comparisons, the Open ASR Leaderboard is a good starting point.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles