Rumble-2-Flash released as open 0.5B speech model for 40+ languages

In a Hugging Face blog post, the authors (writing as "we") announce that they are open sourcing Rumble-2-Flash under the Apache 2.0 license. It is a 0.5B-parameter speech recognition model for 40+ languages, with automatic language detection. The authors say it competes with models 3-60x its size. The weights are on the Hub at Reson8/Rumble-2-Flash, and the model is also available through the authors' API.
The post lists five selling points. The model runs anywhere, with native support in Transformers and vLLM and no custom code. It covers 40+ languages. It offers output control: numbers as digits or words, fillers, verbatim repeats and a forced language. It is fast: an hour of audio in about half a second, with over 7,000 RTFx on full batches. And it holds up on real phone calls, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) on a 302-hour call-centre benchmark.
A quickstart shows how to transcribe with Transformers: install torch, transformers and accelerate, load the processor and AutoModelForMultimodalLM from "Reson8/Rumble-2-Flash", pass audio plus an optional language (omit it to detect automatically), then generate and decode. The audio argument accepts a local path, a URL, a NumPy array, or a list of these.
On how it is built: Rumble-2-Flash uses the Qwen3-ASR architecture for day-0 support. If a stack runs Transformers or vLLM, it runs this model, with no forked runtime or trust_remote_code. To find out how small a model could get before accuracy dropped off, the authors distilled Rumble-2, their larger production model, into it with on-policy distillation. The student does not train on transcripts the teacher wrote. It transcribes the audio itself, and the teacher grades that transcript token by token: at every position it gives its own probability for what should come next, and the student learns to match it. Because the student learns from its own outputs, it is corrected on exactly the mistakes it makes, with feedback on every token instead of one score per clip.
On-policy distillation normally requires teacher and student to share a tokenizer, and these two do not: choosing Qwen3-ASR meant taking its vocabulary too. A method called GOLD removes that restriction. It merges tokens where the two tokenizers split the text differently, so both models are compared over the same text. Tokens present in both vocabularies are compared directly, and the rest by their sorted probabilities.
Results. On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The authors say that is better than Meta's omniASR-LLM 7B, a model 15x larger, and ahead of Whisper large-v3. On the English track it averages 6.6% WER.
On speed, throughput depends on how full the batches are. On the larger datasets in HF's Open ASR Leaderboard, where batches stay full, it reaches over 7,000 RTFx, which the post equates to an hour of audio in about half a second.
On phone calls, which are compressed, noisy and often feature very different accents, the test is AppTek's call-centre benchmark: 302 hours of helpline calls, 14 English accents, scored with AppTek's own script. Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%. That is well ahead of Qwen3-ASR 0.6B (8.7%), from the same model family, and close to Qwen3-ASR 1.7B (7.4%), the only model in the comparison that does better.
The authors close by inviting people to use the Discussions tab to share what they have built, what they like, and what they would change in a next version.
Key facts
- Rumble-2-Flash is a 0.5B-parameter speech recognition model for 40+ languages with automatic language detection, released under Apache 2.0; weights are at Reson8/Rumble-2-Flash on the Hub.
- It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM with no custom code, and was distilled from the larger Rumble-2 with on-policy distillation.
- On the Open ASR Leaderboard multilingual track it averages a 6.96% error rate over nine languages, which the authors say beats Meta's omniASR-LLM 7B (15x larger) and Whisper large-v3; on the English track it averages 6.6% WER.
- On AppTek's 302-hour call-centre benchmark it scores 7.86% WER against 7.78% for NVIDIA's Nemotron 3.5 Transcribe (30B); Qwen3-ASR 1.7B, at 7.4%, is the only model in the comparison that does better.
- The authors report over 7,000 RTFx on full batches, about an hour of audio in half a second; throughput depends on how full the batches are.
Why it matters
The claim is that a 0.5B speech model can stay close to far larger ones: the authors say it competes with models 3-60x its size, and the headline comparison is a 30B NVIDIA model on phone-call audio, where the two differ by 0.08 percentage points of WER (7.86% against 7.78%). The release is under Apache 2.0, with weights on the Hub. The post also describes a training method: on-policy distillation, where the student transcribes audio itself and the teacher grades each token, extended by GOLD to work when teacher and student use different tokenizers.
Who it affects
Developers who already run Transformers or vLLM can use the model as is, since it reuses the Qwen3-ASR architecture and needs no forked runtime or trust_remote_code. Teams working with noisy phone audio, such as helpline calls with varied accents, are the group the call-centre benchmark speaks to. Anyone distilling models across different tokenizers may find the GOLD approach relevant.
How to use it
Load "Reson8/Rumble-2-Flash" with AutoProcessor and AutoModelForMultimodalLM in Transformers, build a request with processor.apply_transcription_request, call generate, and decode with return_format="transcription_only" (or "parsed" if detecting language). Audio can be a local path, a URL, a NumPy array, or a list of these. Pass a language such as "English" to force it, or omit it to detect automatically. The model also offers output control for numbers as digits or words, fillers and verbatim repeats. The weights are Apache 2.0, and the authors also offer an API; no API pricing is given.
How solid is it
The post is first-party: it comes from the model's authors, and the speed and accuracy results are reported by them, with no independent evaluation cited. The numbers are specific: 6.96% average error rate on the Open ASR Leaderboard multilingual track, 6.6% WER on the English track, and 7.86% WER on AppTek's 302-hour call-centre benchmark, scored with AppTek's own script. Weights and code access are public, so the claims can be checked by anyone who runs the model.
Risks and caveats
The multilingual comparison against omniASR-LLM 7B and Whisper large-v3 gives no figures for those two models. Only nine languages are named, as the multilingual leaderboard track; the full list of 40+ is not given. The throughput figure of over 7,000 RTFx applies to full batches and depends on how full they are; the hardware used for the speed measurement is not stated. On the call-centre benchmark the model is not the best: Qwen3-ASR 1.7B scores 7.4% against its 7.86%. The post gives no training data, compute or teacher size.
“The student doesn't train on transcripts the teacher wrote.”
— Rumble-2-Flash blog post, Hugging Face