AI Weekly Malaysia

Back to items Summaries

Rumble-2-Flash: an open 0.5B speech model for 40+ languages

ID
47861
Status
summarized
Published
18 Nov 2026, 8:00 AM
Fetched
18 Nov 2026, 11:50 PM
Provider
Hugging Face Blog
Category
developer-ai
Original URL
https://huggingface.co/blog/rumble-2-flash
Source URL
https://huggingface.co/blog/feed.xml

Summary

Score
7.0
Created
18 Nov 2026, 11:51 PM
Tags
Audience
developersai_ml_learnersai_agent_userssaas_startup_founders

What happened

Reson8 open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model under Apache 2.0 that supports 40+ languages with automatic language detection. It runs natively in Transformers and vLLM, claims about 0.5 seconds to transcribe an hour of audio (7,000+ RTFx on full batches), and matches NVIDIA's 30B Nemotron 3.5 Transcribe on a 302-hour call-centre benchmark. The model uses the Qwen3-ASR architecture and was distilled from Reson8's larger Rumble-2 via on-policy distillation with a GOLD tokenizer-merging step.

Why it matters

Builders evaluating ASR can test a 0.5B Apache-2.0 model that may cut inference cost and latency versus 3-60x larger models, especially for batch call transcription or voice-agent prototypes. The benchmark is vendor-reported, so validate on your own audio before replacing an existing pipeline; the practical decision is whether native Transformers/vLLM support and the speed claim outweigh accuracy gaps in your domain.

Discussion angle

The vendor claims parity with a 30B model on call-centre audio; what failure modes would you probe first - code-switching, background noise, domain jargon, or long silences - before trusting a 0.5B model in production?

Top