Rumble-2-Flash: Open-Source 0.5B Speech Model for 40+ Languages
Summarized by AI from reporting by Hugging Face Blog, published under our editorial policy.
Rumble-2-Flash is a 0.5-billion-parameter speech recognition model that supports over 40 languages. It competes with models 3-60 times its size and is now open-source under the Apache 2.0 license.

Key takeaways
- Rumble-2-Flash is a 0.5-billion-parameter speech recognition model that supports over 40 languages.
- It competes with models 3-60 times its size and is open-source under the Apache 2.0 license.
- The model processes an hour of audio in about half a second with over 7,000 RTFx on full batches.
- Rumble-2-Flash outperforms Meta's omniASR-LLM 7B and Whisper large-v3 on the Open ASR Leaderboard.
- It matches the performance of NVIDIA's Nemotron 3.5 Transcribe (30B) on real phone calls.
Hugging Face released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model that supports over 40 languages. The model is open-source under the Apache 2.0 license and competes with models 3-60 times its size.
How Rumble-2-Flash Works
Rumble-2-Flash runs anywhere with native support in Transformers and vLLM, requiring no custom code. It features automatic language detection and offers output control options such as numbers as digits or words, fillers, verbatim repeats, and forced language. The model processes an hour of audio in about half a second with over 7,000 RTFx on full batches. It also performs well on real phone calls, matching the performance of NVIDIA's Nemotron 3.5 Transcribe, a 30B-parameter model.
The model accepts audio inputs as a local path, a URL, a NumPy array, or a list of these. Rumble-2-Flash uses the Qwen3-ASR architecture, ensuring day-0 support if your stack runs Transformers or vLLM, without the need for a forked runtime or trust_remote_code.
Rumble-2-Flash was created by distilling Rumble-2, a larger production model, using on-policy distillation. In this method, the student model transcribes audio itself, and the teacher grades the transcript token by token. At every position, the teacher gives its own probability for what should come next, and the student learns to match it. Because the student learns from its own outputs, it gets corrected on exactly the mistakes it makes, with feedback on every token instead of one score per clip.
On-policy distillation normally requires teacher and student to share a tokenizer, but these models do not: choosing Qwen3-ASR meant taking its vocabulary too. GOLD removes that restriction by merging tokens where the two tokenizers split the text differently, so both models are compared over the same text. Tokens that exist in both vocabularies are compared directly, and the rest by their sorted probabilities.
Performance and Benchmarks
On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English. That is better than Meta's omniASR-LLM 7B, a model 15 times larger, and ahead of Whisper large-v3. On the English track, it averages 6.6% WER.
Rumble-2-Flash also excels in real-world scenarios. On AppTek's call-centre benchmark (302 hours of helpline calls, 14 English accents), it achieves a 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%. That is well ahead of Qwen3-ASR 0.6B (8.7%), from the same model family, and close to Qwen3-ASR 1.7B (7.4%), the only model in the comparison that does better.
Why It Matters for Everyday Users
Rumble-2-Flash's ability to process speech in multiple languages quickly and accurately makes it a valuable tool for a wide range of applications. From transcribing phone calls to converting speech to text in real-time, this model can enhance productivity and accessibility. Its open-source nature allows developers to integrate it into various applications, making advanced speech recognition technology more accessible to everyone.
How to Get Started
To get started with Rumble-2-Flash, you can access the weights on the Hub at Reson8/Rumble-2-Flash. It is also available through Hugging Face's API. You can experiment with the model by integrating it into your projects or using it to transcribe audio files. Share your creations and feedback in the Discussions tab on the Hugging Face Hub.
Frequently asked
- Is Rumble-2-Flash free to use?
- Yes, Rumble-2-Flash is open-source under the Apache 2.0 license and can be accessed for free.
- What languages does Rumble-2-Flash support?
- Rumble-2-Flash supports over 40 languages, including Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English.
- How fast is Rumble-2-Flash?
- Rumble-2-Flash processes an hour of audio in about half a second with over 7,000 RTFx on full batches.
- Where can I find the weights for Rumble-2-Flash?
- The weights for Rumble-2-Flash are available on the Hub at Reson8/Rumble-2-Flash.