🔍 Read the full analysis: Why Rumble-2-Flash Takes An Open Approach To Speech Across 40+ Languages 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 results on selected multilingual and call-centre tests, but independent evaluations and results across all supported languages have not been provided.
Hugging Face has released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model supporting more than 40 languages, under the Apache 2.0 license. The release makes the model weights available to developers and includes support for Transformers and vLLM; Hugging Face says its results on selected benchmarks put the compact model near or ahead of several larger systems.
The model is listed on Hugging Face as Reson8/Rumble-2-Flash and is also accessible through Hugging Face’s API. According to the company, it can automatically detect a language or use one specified by the developer. Transcription controls let users choose how numbers are rendered and whether fillers and verbatim repetitions are retained. The release also includes a Transformers quickstart and accepts audio from a local path, URL, NumPy array or list of inputs.
Hugging Face reports a 6.96% average error rate on the multilingual track of the Open ASR Leaderboard. That test covers Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The company says the score is better than results for Meta’s 7-billion-parameter omniASR-LLM and Whisper large-v3. On the leaderboard’s English track, Hugging Face reports a 6.6% average word error rate.
For a separate phone-call evaluation, Hugging Face says the model recorded a 7.86% word error rate on AppTek’s 302-hour call-centre benchmark, which includes 14 English accents. The company describes that as on par with NVIDIA Nemotron 3.5 Transcribe, reported at 7.78% in the comparison. It lists Qwen3-ASR 0.6B at 8.7% and Qwen3-ASR 1.7B at 7.4%. Hugging Face says the test was scored using AppTek’s script.
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
Rumble-2-Flash’s reported results put a 0.5-billion-parameter model in comparisons with substantially larger speech systems. If those results hold for a developer’s own data and languages, a smaller model could make transcription more practical for teams with limited computing capacity. The Apache 2.0 license also gives developers broad permission to inspect, adapt and deploy the released weights, subject to the license terms.
The call-centre result is relevant because telephone audio can include background noise, compression and varied accents. However, a score on one named benchmark does not establish performance on other call-centre recordings or prove that the model matches larger systems across all conditions. Hugging Face says the model works with the widely used Transformers and vLLM stacks without a custom runtime or trust_remote_code, which may ease adoption for teams already using those tools.
The company also reports throughput above 7,000 RTFx on full batches in larger leaderboard datasets, describing this as processing an hour of audio in about half a second. That is a company-reported measurement, not a general speed guarantee: Hugging Face says throughput depends on batch size, while actual results will also depend on hardware and deployment setup.
Products Worth Considering

Samsung 32 Odyssey G5 QHD Gaming Monitor

Logitech MX Creative Console with 9 Customizable Keys, Graphite

Apple Watch SE 3 GPS + Cellular 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How Hugging Face Built the Model
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 goal as finding how small the model could become before accuracy declined substantially. In its account of the process, the smaller model transcribes audio while the larger model evaluates the student’s next-token probabilities and provides feedback at each position.
The company calls its method GOLD. Hugging Face says it addresses a tokenizer mismatch between teacher and student models by merging tokens when the systems divide text differently, allowing their outputs to be compared over the same text. Shared vocabulary tokens are compared directly; remaining tokens are compared by their sorted probabilities. These technical details and the performance figures come from Hugging Face’s release information.
““an hour of audio in about half a second””
— Hugging Face
Benchmark Results Need Independent Checks
The supplied release information does not include independent evaluations confirming the leaderboard, throughput or call-centre results. The multilingual leaderboard figure covers nine named languages, not every language in the model’s stated support of more than 40. Results may vary with accents, recording conditions, transcription settings, scoring methods and the audio used.
The reported speed applies to full batches, and the announcement does not establish how quickly the model handles individual requests or different hardware configurations. It also does not show how accuracy changes when users adjust the transcription controls. These limits mean the figures should be treated as results for specified tests, not as a guarantee of performance in every deployment.
Testing Across Real-World Audio
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 share feedback and examples in the model’s Discussions tab, including requests for changes in a future version.
The next evidence to watch for is testing by developers and independent evaluators on additional languages, accents and recording environments, along with speed measurements on the hardware and batch sizes used in practice. Hugging Face did not state a timetable for further releases or independent tests.
Key Questions
What is Rumble-2-Flash?
It is a 0.5-billion-parameter speech recognition model that Hugging Face says supports more than 40 languages. Its weights are available under the Apache 2.0 license.
Where can developers access the model?
The weights are listed as Reson8/Rumble-2-Flash on Hugging Face. The company also says the model is available through its API and supports Transformers and vLLM.
How well did it perform on the reported tests?
Hugging Face reports a 6.96% average error rate on a multilingual leaderboard covering nine languages, a 6.6% average word error rate on its English track, and a 7.86% word error rate on AppTek’s call-centre benchmark. These are company-reported results; independent confirmation was not included in the supplied information.
Do the benchmarks prove it works equally well in all 40-plus languages?
No. The reported multilingual leaderboard result covers nine named languages. The release does not provide results for every language the model is said to support or for every accent and audio condition.
What affects the model’s reported transcription speed?
Hugging Face says its throughput above 7,000 RTFx was measured on full batches and depends on batch size. Hardware and deployment conditions also affect practical speed, so the figure is not a guarantee for individual requests.
Primary source: Hugging Face · via ThorstenMeyerAI.com
