AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: 5 Things To Know About Rumble-2-Flash And Its 40+ Languages on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Hugging Face has released Rumble-2-Flash, an open-source, 0.5-billion-parameter speech recognition model described as supporting more than 40 languages. The company reports competitive leaderboard and call-centre results, but the figures have not been independently verified and do not establish performance across all languages or deployments.

Hugging Face has open-sourced Rumble-2-Flash, a speech recognition model with 0.5 billion parameters that the company says supports more than 40 languages. The Apache 2.0 release gives developers access to the weights, Transformers and vLLM support, and the option to test the model through Hugging Face’s API; its published performance figures are company evaluations, not independent results.

The model is available on Hugging Face under the identifier Reson8/Rumble-2-Flash. Hugging Face says it can automatically detect a language or accept a language specified by the developer. Transcription options include how numbers are written and whether fillers and verbatim repetitions are retained. The release also includes a Transformers quickstart and supports audio supplied through a local path, a URL, a NumPy array or a list of inputs.

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. On the leaderboard’s English track, Hugging Face reports a 6.6% average word error rate. These scores apply to the named evaluation tracks and do not show results for every language the model supports.

For phone audio, 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. Qwen3-ASR 0.6B scored 8.7%, while Qwen3-ASR 1.7B scored 7.4%. Hugging Face says the test was scored with AppTek’s script.

At a glance
announcementWhen: Announced; model is available now
The developmentHugging Face released Rumble-2-Flash under the Apache 2.0 license, giving developers access to its model weights and speech recognition tools.
Speech AI · Briefing

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.

0.5B
model parameters
40+
supported languages
7.86%
reported phone-call word error rate

What developers get

The release is designed to fit familiar tools and flexible transcription workflows.

Open release

Weights and API

Find the weights at Reson8/Rumble-2-Flash or access the model through Hugging Face’s API.

Integration

Transformers and vLLM

Hugging Face says the model works with both stacks without a custom runtime or trust_remote_code.

Transcription

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.

EvaluationReported resultContext
Multilingual leaderboard6.96% average error rateNine named languages
English leaderboard6.6% average word error rateEnglish track
AppTek call-centre benchmark7.86% word error rate302 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.

1
Student transcribes

The smaller model produces a transcription from the audio.

2
Teacher evaluates

The larger model provides feedback on next-token probabilities.

3
Text is aligned

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.

Potential
  • A compact model may suit tighter compute budgets.
  • Reported call results cover varied English accents.
  • Full-batch throughput is reported above 7,000 RTFx.
Still to check
  • Results are from Hugging Face’s evaluations.
  • Performance across all 40-plus languages is not established.
  • Speed depends on hardware, batch size, and deployment.
Benchmark scopeThe reported half-second per hour of audio applies to full batches in larger leaderboard datasets, not a guaranteed speed for individual requests.
The takeaway
Rumble-2-Flash makes an open, compact speech model available to test; its benchmark claims need validation on your languages, recordings, and hardware.

A Smaller Model for Speech Tasks

Rumble-2-Flash’s size may make it useful to teams working with limited computing resources, if its accuracy proves suitable for their audio and languages. Its support for Transformers and vLLM may also make it easier to test in systems that already use those tools; Hugging Face says it runs without a custom runtime or trust_remote_code.

The call-centre result is relevant because phone recordings can include background noise, compression and varied accents. On this particular test, the reported score puts the compact model near larger systems. That is a reason for developers to evaluate it, not proof that it will match those systems on other call-centre recordings or in production. Hugging Face also reports throughput above 7,000 RTFx on full batches in larger leaderboard datasets, which it describes as processing an hour of audio in about half a second. The company says speed depends on batch size; hardware and deployment setup also affect real-world results.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

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. In the company’s account, the smaller model transcribes audio while the larger model evaluates the student’s next-token probabilities and provides feedback at each position.

The process uses a method Hugging Face calls GOLD, which the company says addresses differences between the teacher and student tokenizers. When the models divide text into tokens differently, the method merges tokens so their outputs can be compared over the same text. Hugging Face says shared vocabulary tokens are compared directly and other tokens are compared by their sorted probabilities. The model’s weights and implementation support are available with the release; the supplied announcement does not include an independent review of the training method or its reported scores.

““an hour of audio in about half a second””

— Hugging Face

Where the Scores May Not Apply

The reported leaderboard, phone-call and throughput results come from Hugging Face’s own evaluations; the source material provides no independent confirmation. The tests cover specific datasets, languages, accents and scoring conditions, so they do not establish how the model performs across all 40-plus supported languages or on other recording environments.

It is also unclear how transcription accuracy changes with the model’s style controls or across different real-world workloads. The fast throughput figure is for full batches, not a guarantee for individual requests. Results may vary with hardware, batch size and deployment configuration, and the announcement does not provide a single measure covering those variations.

Developers Can Test the Release

Developers can download the weights from Reson8/Rumble-2-Flash or access the model through Hugging Face’s API, then compare it against their own audio, languages and hardware. Hugging Face has invited users to share feedback and examples in the model’s Discussions tab, including suggestions for future changes.

Further evidence will depend on testing outside the company’s announcement, particularly evaluations across additional languages and recording conditions, and speed measurements on the batch sizes and hardware developers use. The supplied release information does not state a schedule for independent testing or a future version.

Key Questions

What is Rumble-2-Flash?

It is a 0.5-billion-parameter speech recognition model released by Hugging Face. The company says it supports more than 40 languages and provides automatic language detection and transcription controls.

Is Rumble-2-Flash open source?

Hugging Face released the model under the Apache 2.0 license. Its weights are available at Reson8/Rumble-2-Flash, and the company says it can also be accessed through its API.

How accurate is the model?

Hugging Face reports a 6.96% average error rate on the multilingual Open ASR Leaderboard track and a 7.86% word error rate on AppTek’s 302-hour call-centre benchmark. These are company-reported results for specific tests, not independent confirmation or a guarantee for other audio.

Does the model support all languages equally well?

The release says the model supports more than 40 languages, but the cited leaderboard result covers nine named languages. The supplied material does not give benchmark scores for every supported language, so equal performance across them has not been established.

How fast is Rumble-2-Flash?

Hugging Face reports throughput above 7,000 RTFx on full batches in larger leaderboard datasets and describes that as about half a second to process an hour of audio. The company says throughput depends on batch size; actual speed also depends on hardware and deployment setup.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

Who Are The Three People Behind Claude Code Leaving Anthropic?

Analytics India Magazine reports three people linked to Claude Code left Anthropic within 10 days, but their identities and reasons are unknown.

Phonon TTS Launches With High-Quality On-Device Speech

Hugging Face’s headline presents Phonon TTS as on-device speech that tops benchmarks, but available material includes no results or release details.

Making Advanced AI Models More Cost-Efficient: Claude Opus 5.5

Anthropic’s Claude Opus 5.5 reduces operational costs by 40%, improves speed, and enhances efficiency for AI-driven tasks, challenging competitors.

AI Signal Monitor: Born Against And The Hobby Coding Discussion

IdeaNavigator AI proposes testing a role-filtered monitor for AI developments, using a Hacker News discussion about LLM use as a pilot item.