Reson8 has released Rumble-2-Flash, a roughly 500-million-parameter automatic speech recognition model, as downloadable weights on Hugging Face. The November 18 release declares Apache-2.0 licensing and targets transcription in more than 40 languages. Unlike the company's separately offered hosted speech products, the newly shared checkpoint can be inspected and run by developers who provide their own inference environment. That makes the release relevant to teams building private transcription pipelines, not merely customers choosing another API. Primary source ↗
The technical choice is deliberately familiar: the model uses the Qwen3-ASR architecture and is documented for direct loading in Hugging Face Transformers and vLLM. Reson8 says users can let the model detect the language automatically or influence formatting, including how numbers, filler words and repeated speech appear in transcripts. These controls matter in real applications, where a clean-looking sentence can conceal a missed amount, name or hesitation.
The team reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track covering Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The language-level metric is not identical everywhere: Chinese is reported using character error rate, while word error rate is standard for the other comparisons. Reson8 also cites a 302-hour, 14-accent AppTek call-center evaluation, giving Rumble-2-Flash 7.86% WER versus 7.78% for NVIDIA's much larger Nemotron 3.5 Transcribe. These are release-author comparisons, not an independent Open Model Weights result.
The publisher's headline throughput claim is likewise workload-specific. Processing an hour of audio in roughly half a second and exceeding 7,000 times real-time speed refers to heavily batched inference under the reported serving setup, including an NVIDIA H200. It should not be interpreted as the latency of one recording on a laptop or as proof of cheap real-time streaming for every device.
For deployment, the interesting questions are repeatability and trust. A practical evaluation should pin the exact model repository revision, verify the license and supporting artifacts, and test accents, code switching, noisy calls, punctuation, privacy handling and long recordings. The availability of source weights does not automatically validate performance or any particular data-governance configuration.
Rumble-2-Flash is a promising new small speech-model candidate for the registry, but today its strongest public claims remain publisher supplied. Open Model Weights should only convert this news into a model profile once the exact downloadable files, source revision and licensing evidence have passed the usual independent artifact checks.
What this changes for the evidence layer
New public checkpoint candidate, not yet a verified catalog entry. Pin the official Reson8/Rumble-2-Flash SHA, audit exact weight artifacts and Apache-2.0 license evidence, CTC files and inference compatibility before adding a source-verified model profile. Publisher-reported WER, CER and throughput are not Open Model Weights benchmarks.
Reson8 — original Rumble-2-Flash Hugging Face announcement and model card
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗