Small Rumble-2-Flash speech model for many languages
Hugging Face has released the Rumble-2-Flash speech model under an open licence, and it supports many languages.

Key takeaways
- The model uses a 0.5B parameter design and supports more than 40 languages.
- It reached 7.86% error rate on a 302-hour call centre benchmark.
- It performed better than the Meta model on the Open ASR Leaderboard.
- It runs through Hugging Face Transformers and vLLM without extra work.
In this article
What changed
The Rumble-2-Flash speech model was released on 18 November 2026 by Hugging Face. The system is an open speech recognition model. It has 0.5B parameters and supports more than 40 languages. The company said it can compete with larger models.
Those models were three to 60 times bigger. The model arrived under the Apache 2.0 licence.
How it performs
The Qwen3-ASR design sits at the heart of Rumble-2-Flash. Hugging Face said it gained support on the same day. The model can run through Hugging Face Transformers and vLLM. It needs no forked runtime or special remote code.
The team reported results on a 302-hour call centre benchmark. The model reached 7.86% error rate. A 30B model from NVIDIA reached 7.78%. The smaller model was close.
It was ahead of Qwen3-ASR 0.6B. That model had an 8.7% error rate. It was close to Qwen3-ASR 1.7B. That model had a 7.4% error rate.
On the Open ASR Leaderboard, its average error rate was 6.96% across several languages. That was better than the Meta model. The Meta model is 15 times larger. It was ahead of Whisper large-v3.
The model was fast. On full batches, it reached over 7,000 RTFx. An hour of audio took about 0.5 seconds. It offers language detection and output controls.
The team used on-policy distillation. The method let the smaller model learn from its own outputs. The teacher gave feedback on each token. The team used a method called GOLD to merge tokens.
Who it affects

The model may affect small teams and independent researchers. It offers speech recognition for more than 40 languages. The list includes Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. It also covers other languages.
The automatic language detection may help when the source language is not known. The output controls may suit noisy calls. The model may be useful for voice apps, call centres and research.
What it means next
The release is open, so others can inspect and build on it. Hugging Face placed the weights on its Hub. The model is also available through its API.
The company asked users to share what they create. It said it is curious about new uses. The next version may change things.
Sources
- Rumble-2-Flash: an open 0.5B speech model for 40+ languagesHugging Face Blog
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



