🔍 Read the full analysis: Could Rumble-2-Flash Make Multilingual Speech AI More Accessible? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Hugging Face has released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model for more than 40 languages, under the Apache 2.0 license. The company reports competitive leaderboard and call-centre results, but independent tests are not included and real-world performance will depend on language, audio, hardware and workload.
Hugging Face has open-sourced Rumble-2-Flash, a 0.5-billion-parameter speech recognition model designed for more than 40 languages. Its Apache 2.0 license, published weights and support for common inference tools give developers a way to test the system, while the performance figures released so far come from Hugging Face’s own evaluations.
The model is available on Hugging Face as Reson8/Rumble-2-Flash and through Hugging Face’s API. The company says it supports automatic language detection and transcription options for handling numbers, fillers and repeated words. Developers can specify a language or leave it for the model to detect. The release also includes support for Transformers and vLLM, without requiring a custom runtime or trust_remote_code, according to the announcement.
Hugging Face reports a 6.96% average error rate on the multilingual track of the Open ASR Leaderboard. That track covers Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The company says the result is better than Meta’s 7-billion-parameter omniASR-LLM and Whisper large-v3. For the leaderboard’s English track, it reports a 6.6% average word error rate.
On AppTek’s 302-hour call-centre benchmark, which includes 14 English accents, Hugging Face reports a 7.86% word error rate for Rumble-2-Flash. It describes that as close to NVIDIA Nemotron 3.5 Transcribe’s 7.78% score. The announcement lists Qwen3-ASR 0.6B at 8.7% and Qwen3-ASR 1.7B at 7.4%. Hugging Face says it used AppTek’s scoring script for this comparison; the results are tied to that test and do not establish performance on other call-centre recordings.
Rumble-2-Flash Open speech recognition
Hugging Face released a compact speech model with weights under the Apache 2.0 license. Its reported benchmark results are promising, but developers should check them against their own languages and audio.
What developers get
The release is designed to fit familiar tools and flexible transcription workflows.
Weights and API
Find the weights at Reson8/Rumble-2-Flash or access the model through Hugging Face’s API.
Transformers and vLLM
Hugging Face says the model works with both stacks without a custom runtime or trust_remote_code.
Adjustable output
Choose language handling and whether to retain fillers, repetitions, or numbers as digits or words.
Reported benchmark results
Hugging Face reports these scores on named evaluations; they are not independent tests.
| Evaluation | Reported result | Context |
|---|---|---|
| Multilingual leaderboard | 6.96% average error rate | Nine named languages |
| English leaderboard | 6.6% average word error rate | English track |
| AppTek call-centre benchmark | 7.86% word error rate | 302 hours; 14 English accents |
How the model was made
Hugging Face says it distilled its larger Rumble-2 production model using the Qwen3-ASR architecture.
The smaller model produces a transcription from the audio.
The larger model provides feedback on next-token probabilities.
The GOLD method addresses differences in how the models tokenize text.
What the claims mean in practice
The announcement points to potential efficiency, with conditions developers should account for.
- A compact model may suit tighter compute budgets.
- Reported call results cover varied English accents.
- Full-batch throughput is reported above 7,000 RTFx.
- Results are from Hugging Face’s evaluations.
- Performance across all 40-plus languages is not established.
- Speed depends on hardware, batch size, and deployment.
A Smaller Model for Speech Workloads
A compact speech model could be useful to developers working with limited compute budgets, particularly if it performs well on the languages and audio they handle. A 0.5-billion-parameter system may be easier to test and deploy than much larger models, although actual costs and performance depend on hardware and serving setup. The Apache 2.0 license also allows developers to inspect and adapt the released weights under that license, rather than relying only on a hosted service.
The call-centre result gives the release a practical point of comparison: phone audio can include compression, background noise and accent variation. Still, a score on one named benchmark is not evidence that the model will match larger systems across different customers, languages or recording conditions. For teams considering adoption, their own evaluation data will matter more than a single published ranking.
Hugging Face also reports throughput above 7,000 RTFx on full batches in larger leaderboard datasets, describing it as processing an hour of audio in about half a second. The company says speed depends on batch size. That figure is therefore not a general estimate for individual requests or every deployment: hardware, batch fullness and configuration can change the result.
Products Worth Considering

DJI Mini 4K Drone with 4K UHD Camera, Under 249 g

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger

Apple Watch Series 12 GPS 42mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How Rumble-2-Flash Was Built
Hugging Face says Rumble-2-Flash uses the Qwen3-ASR architecture and was created by distilling its larger Rumble-2 production model. The company describes the effort as a way to find how much the model could be reduced before accuracy fell substantially. In the process it outlines, the smaller model transcribes audio, while the larger model assesses the student’s next-token probabilities and provides feedback at each position.
The company calls its distillation method GOLD. According to Hugging Face, the method addresses differences in how the teacher and student tokenizers divide text. It combines tokens where needed so the systems’ outputs can be compared over the same text, comparing shared vocabulary tokens directly and handling remaining tokens through sorted probabilities. The release provides a quickstart and accepts audio from a local path, URL, NumPy array or list of inputs.
The announcement’s leaderboard, phone-call and throughput figures, along with its account of the training method, are Hugging Face’s claims. The supplied release information does not include independent evaluations. Its reported multilingual score covers nine named languages, while the model’s stated support extends to more than 40; those are different scopes.
““an hour of audio in about half a second””
— Hugging Face
Independent Tests Still Needed
No independent evaluation is included in the release material, so the reported accuracy and speed figures have not been independently confirmed here. The announcement does not show how the model performs on every language among the 40-plus it supports, or across accents, microphone quality, background noise and recording environments beyond the named tests.
It is also not clear how its transcription options affect accuracy, or whether the reported results will hold for a developer’s particular workload. The high throughput figure applies to full batches on specified leaderboard datasets, not necessarily to low-volume or single-request use. The available information gives no general hardware configuration or independent schedule for further testing.
Developers Can Test the Release
Developers can download the weights from Reson8/Rumble-2-Flash or try the model through Hugging Face’s API. The company has invited users to post examples and feedback in the model’s Discussions tab, including suggestions for future changes. No further release date or independent evaluation schedule was stated.
The next evidence to watch for is testing by outside users on additional languages, varied audio and deployment hardware. Results that report the data, scoring method, batch size and hardware would help teams compare the company’s figures with performance in their own settings.
Key Questions
What is Rumble-2-Flash?
It is a 0.5-billion-parameter speech recognition model from Hugging Face, with stated support for more than 40 languages. Its weights are available under the Apache 2.0 license.
Where can developers access the model?
The weights are published on Hugging Face under Reson8/Rumble-2-Flash. Hugging Face also says the model is available through its API and works with Transformers and vLLM.
Has its performance been independently verified?
The cited leaderboard, call-centre and throughput results are reported by Hugging Face. The announcement does not provide independent evaluations, so developers should test the model on relevant audio and languages.
Does the model process an hour of audio in half a second?
Hugging Face reports that speed for full batches on larger leaderboard datasets and says throughput depends on batch size. It is not a general guarantee for individual requests or different hardware.
Primary source: Hugging Face · via ThorstenMeyerAI.com
