TL;DR
Hugging Face has open-sourced 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 strong leaderboard and phone-call benchmark results, but those figures are its own evaluations and do not establish performance across every language or deployment.
Hugging Face has open-sourced Rumble-2-Flash, a 0.5-billion-parameter speech recognition model for more than 40 languages, under the Apache 2.0 license. The release gives developers access to model weights and support in Transformers and vLLM; Hugging Face says its benchmark results put the compact system close to, or ahead of, several much larger speech models.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The model is available on Hugging Face under the identifier Reson8/Rumble-2-Flash. The company says it supports automatic language detection and controls for transcription style, including whether numbers appear as digits or words, whether fillers and verbatim repetitions are retained, and whether a language is specified. It is also available through Hugging Face’s API.
Hugging Face reports a 6.96% average error rate on the multilingual track of the Open ASR Leaderboard, measured across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. It says that result beats Meta’s 7-billion-parameter omniASR-LLM and Whisper large-v3. On the English track, the company reports an average word error rate of 6.6%.
For a phone-call evaluation, Hugging Face says Rumble-2-Flash scored 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, which scored 7.78% in the comparison. Qwen3-ASR 0.6B scored 8.7%, while Qwen3-ASR 1.7B scored 7.4%. Hugging Face says the benchmark was scored with AppTek’s own 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.
Small Models, Broader Speech Access
A model with half a billion parameters could make speech transcription easier to run in settings with tighter compute budgets, if the reported accuracy holds for a developer’s languages, audio and workload. Hugging Face says the model works with widely used Transformers and vLLM stacks without a custom runtime or trust_remote_code, which may reduce integration friction for teams already using those tools.
The reported phone-call result is relevant because compressed, noisy calls and varied accents can challenge transcription systems more than clean recordings. The comparison suggests that a compact model can approach the score of a much larger model on this particular benchmark. It does not establish that the smaller model performs equally well across other call-centre data, languages, accents or scoring methods.
Hugging Face also reports throughput above 7,000 RTFx on full batches in larger Open ASR Leaderboard datasets, describing that as processing an hour of audio in about half a second. The company cautions that throughput depends on batch size. That qualification matters for practical use: real-world speed can differ with hardware, batch fullness and deployment setup.
Products Worth Considering

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite

Bose QuietComfort Ultra Wireless Earbuds (2nd Gen) with Deep Plum Finish and 20W Dual-Port Charger

LG 27GR83Q-B 27-inch Ultragear QHD IPS Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Distillation Behind the Release
Hugging Face says Rumble-2-Flash uses the Qwen3-ASR architecture and was made by distilling its larger Rumble-2 production model. The stated goal was to find how small the model could become before its accuracy fell off. In the company’s description of the process, the smaller model transcribes audio itself, then the larger model evaluates the student’s next-token probabilities and provides feedback at each position.
The company says its method, called GOLD, addresses a tokenizer mismatch between the teacher and student models. It merges tokens when the two systems divide text differently so their outputs can be compared over the same text. Hugging Face says shared vocabulary tokens are compared directly and remaining tokens by their sorted probabilities.
The release includes a quickstart for Transformers and supports audio from a local path, URL, NumPy array or list of those inputs. Developers can provide a language or leave it unspecified for automatic detection. The benchmark claims and implementation details in the announcement come from Hugging Face; the supplied information does not include independent evaluations.
““an hour of audio in about half a second””
— Hugging Face
Benchmark Scope and Runtime Limits
The announcement does not provide independent confirmation of the reported leaderboard, speed or call-centre results. The figures are tied to named tests and evaluation conditions, and do not show how the model performs on every language among the stated 40-plus, on other accents or on audio from different recording environments.
Practical throughput will depend on deployment details, including hardware and batch size. The reported full-batch speed is not a general guarantee for individual requests. The release information also does not establish how accuracy changes with the model’s transcription controls or across varied real-world workloads.
Developers Can Test the Weights
The model weights are available at Reson8/Rumble-2-Flash, and Hugging Face says developers can also access the model through its 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 useful evidence will come from evaluations outside the company’s announcement: results on additional languages and audio conditions, and speed measurements on the hardware and batch sizes developers actually use. No further release date or independent test schedule was stated.
Where I land
My assessment is that Rumble-2-Flash is a useful release to watch because its reported results put a relatively small model in the same conversation as much larger systems, and the Apache 2.0 license gives developers room to evaluate it. The phone-call test makes the announcement more relevant than a speed claim alone, since it addresses noisy, accented audio.
The strongest counterargument is that the key numbers are presented by the model’s maker and depend on defined benchmarks and batch conditions. Those results do not yet show that a team will get comparable accuracy or speed on its own recordings. I would put more weight on the claims after independent tests across more languages, hardware configurations and real deployments reproduce them.
Source: Hugging Face
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 and automatic language detection.
Is the model open source?
Hugging Face released it under the Apache 2.0 license. Its weights are listed at Reson8/Rumble-2-Flash.
How did it perform on phone calls?
Hugging Face reports a 7.86% word error rate on AppTek’s 302-hour call-centre benchmark across 14 English accents. The company compared this with 7.78% for NVIDIA Nemotron 3.5 Transcribe; the figures are specific to that benchmark.
How fast is transcription?
The company reports more than 7,000 RTFx on full batches in larger leaderboard datasets, or about half a second to process an hour of audio. It says throughput depends on batch fullness, so this is not a universal per-request speed.
Source: Hugging Face
