Reson8, Mark 't Hart, Eustache Le Bihan, Eric Bezzam · Nov 18, 2026
Hugging Face article presents Reson8’s Rumble-2-Flash speech model
Markdown · JSON · Original source
’s article announces Reson8’s open speech recognition model Rumble-2-Flash, with 0.5B parameters and support for more than 40 languages. Reson8 says it competes with models 3–60x its size and runs in Transformers and vLLM without custom code. The team says it used on-policy distillation to make a smaller model: the student transcribes audio, then learns from a larger model’s token-by-token feedback. GOLD handles differences between the models’ . On the Open ASR Leaderboard multilingual track, Reson8 reports a 6.96% average error rate across nine languages; on the English track, it reports 6.6% WER. On AppTek’s of 302 hours of helpline calls in 14 English accents, the model scored 7.86% WER, compared with 7.78% for ’s Nemotron 3.5 Transcribe. Reson8 also reports throughput above 7,000 RTFx on larger leaderboard datasets with full batches, equivalent to about half a second for an hour of audio. The article presents these as benchmark results; it does not establish performance on other audio or in other deployment settings.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Nov 18, 2026 (2026-11-18T00:00:00.772Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-18T19:34:19.294Z. First observed: 2026-11-18T14:43:04.732Z. Retrieved: 2026-11-18T19:33:07.792Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The performance figures are claims reported in the post and are tied to the named and conditions; they do not establish performance on other audio or in other deployment settings. The post reports throughput for larger leaderboard datasets where batches stay full, so that is not presented as a result for every workload.
Open question
How does Rumble-2-Flash perform on audio and in deployment settings beyond the described in the article?
Source-stated
Reson8 announced Rumble-2-Flash as an open speech recognition model with 0.5B parameters and support for more than 40 languages.
Today we're open sourcing Rumble-2-Flash under the Apache 2.0 license. It's a 0.5B-parameter speech recognition model for 40+ languages
Source-stated
Reson8 says the model was trained with on-policy distillation, in which the student transcribes audio and receives token-by-token feedback from the teacher.
The student doesn't train on transcripts the teacher wrote. It transcribes the audio itself, and the teacher grades that transcript token by token
Source-stated
The post reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track across nine named languages, and 6.6% WER on its English track.
Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That's better than Meta's omniASR-LLM 7B
Source-stated
On AppTek’s call-centre benchmark, Reson8 reports 7.86% WER for Rumble-2-Flash and 7.78% for NVIDIA’s Nemotron 3.5 Transcribe.
Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%.
Source-stated
Reson8 reports throughput above 7,000 RTFx on larger leaderboard datasets when batches stay full.
on the larger datasets in HF's Open ASR Leaderboard, where batches stay full, it reaches over 7,000 RTFx.
Source and contributor context
Reson8, Mark 't Hart, Eustache Le Bihan, Eric Bezzam · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Model tokens
Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Figure
Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Editorial provenance
Version 1. Publication ID: c5432fb2-f7f2-4d3e-8577-011c3a71fcc4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T19:34:19.294Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 9af96bcd1e8c2caa39933ac19f0102ef7576081e66b20658eb6de013b827ff6c. Source text hash: 534bf6bbac111d013823ea644911face72ac5c197e68c86057de56a40d213234.
Revision and source-check history
Version 1 · 2026-11-18T19:34:19.294Z
Hugging Face article presents Reson8’s Rumble-2-Flash speech model
’s article announces Reson8’s open speech recognition model Rumble-2-Flash, with 0.5B parameters and support for more than 40 languages. Reson8 says it competes with models 3–60x its size and runs in Transformers and vLLM without custom code. The team says it used on-policy distillation to make a smaller model: the student transcribes audio, then learns from a larger model’s token-by-token feedback. GOLD handles differences between the models’ . On the Open ASR Leaderboard multilingual track, Reson8 reports a 6.96% average error rate across nine languages; on the English track, it reports 6.6% WER. On AppTek’s of 302 hours of helpline calls in 14 English accents, the model scored 7.86% WER, compared with 7.78% for ’s Nemotron 3.5 Transcribe. Reson8 also reports throughput above 7,000 RTFx on larger leaderboard datasets with full batches, equivalent to about half a second for an hour of audio. The article presents these as benchmark results; it does not establish performance on other audio or in other deployment settings.
Original source
· Article
Source details
Original source date: Nov 18, 2026 (2026-11-18T00:00:00.772Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-18T19:34:19.294Z. First observed: 2026-11-18T14:43:04.732Z. Retrieved: 2026-11-18T19:33:07.792Z.
Limits
The performance figures are claims reported in the post and are tied to the named and conditions; they do not establish performance on other audio or in other deployment settings. The post reports throughput for larger leaderboard datasets where batches stay full, so that is not presented as a result for every workload.
Open question
How does Rumble-2-Flash perform on audio and in deployment settings beyond the described in the article?
Source-stated
Reson8 announced Rumble-2-Flash as an open speech recognition model with 0.5B parameters and support for more than 40 languages.
Today we're open sourcing Rumble-2-Flash under the Apache 2.0 license. It's a 0.5B-parameter speech recognition model for 40+ languages
Source-stated
Reson8 says the model was trained with on-policy distillation, in which the student transcribes audio and receives token-by-token feedback from the teacher.
The student doesn't train on transcripts the teacher wrote. It transcribes the audio itself, and the teacher grades that transcript token by token
Source-stated
The post reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track across nine named languages, and 6.6% WER on its English track.
Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That's better than Meta's omniASR-LLM 7B
Source-stated
On AppTek’s call-centre benchmark, Reson8 reports 7.86% WER for Rumble-2-Flash and 7.78% for NVIDIA’s Nemotron 3.5 Transcribe.
Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%.
Source-stated
Reson8 reports throughput above 7,000 RTFx on larger leaderboard datasets when batches stay full.
on the larger datasets in HF's Open ASR Leaderboard, where batches stay full, it reaches over 7,000 RTFx.
Source and contributor context
Reson8, Mark 't Hart, Eustache Le Bihan, Eric Bezzam · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Model tokens
Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Figure
Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Editorial provenance
Version 1. Publication ID: c5432fb2-f7f2-4d3e-8577-011c3a71fcc4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T19:34:19.294Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 9af96bcd1e8c2caa39933ac19f0102ef7576081e66b20658eb6de013b827ff6c. Source text hash: 534bf6bbac111d013823ea644911face72ac5c197e68c86057de56a40d213234.
uncertain · 2026-11-18T19:34:19.294Z · anthropic/claude-opus-5.5
Publication: c5432fb2-f7f2-4d3e-8577-011c3a71fcc4
attribution — pass: The post is a Hugging Face blog article co-authored by Reson8 and Hugging Face contributors. The 'we' announcing the release is the Reson8 team, and the reading attributes performance claims to Reson8.
Today we're open sourcing Rumble-2-Flash under the Apache 2.0 license.
summary — uncertain: The facts are accurate, but a general reader is left with unexplained central jargon. WER is never defined as word error rate (the share of words transcribed wrongly). GOLD, tokenizer, vLLM and Transformers are not explained, and RTFx is only partly explained. The reading also drops the size contrast that gives the call-centre result its force: Nemotron is a 30B model and Rumble-2-Flash is 0.5B. Why the model might matter (small, open, fast) is never stated. Smallest repair: define WER and tokenizer in plain words, note Nemotron's 30B size, and add one source-supported sentence on significance.
on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%
date — pass: The originalPublished date, 2026-11-18, matches the date on the publisher page.
Published November 18, 2026
support — pass: All claims and figures match the source: 0.5B parameters, 40+ languages, 6.96% and 6.6% error rates, 7.86% vs 7.78% on the call-centre benchmark, and over 7,000 RTFx with full batches. The limitations and the question are appropriately conditional and do not invent benefits or costs.
Its throughput depends on how full the batches are
media — pass: The selected media is the verified link to the original article, which is the direct source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceSmall chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceFigure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Explore FigureA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceSmall chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceFigure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Explore FigureA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference