Reson8 announced open weights for Rumble-2-Flash on November 18, 2026. Its Hugging Face model card lists 500 million parameters, 45 languages and Apache 2.0 licensing. SourceSource
The card lists native Hugging Face Transformers and vLLM support, with installation and inference examples for both. Those include local inference: the Transformers example accepts local audio paths, while the vLLM example serves the model through localhost and reads an audio file. Source
The documented transcription options let users:
- Choose digits or words for numbers. Digits are the default.
- Keep fillers such as “um” and “uh.” These are omitted by default.
- Collapse repeated words. Repetitions are retained by default; disabling the
verbatim option collapses them.
- Specify a language or leave it to automatic detection. The language setting takes its English name, such as “Dutch” or “English.” Source
Your hesitation now has a configuration flag. The controls alone don’t show how accurately the model recognizes speech. Source
In the original announcement post, @raoulritterr said Rumble-2-Flash beats models 20 times its size and is small enough to run on a phone. The post names neither comparison models nor phone hardware. The model card supplies no benchmark results, runtime measurements or memory requirements to substantiate those claims. Phone-ready is a claim here, not a hardware recommendation. SourceSource
To try it locally, head to the Hugging Face repository for the weights and setup examples. The captured page lists no deployment through Hugging Face Inference Providers. Source
Comments