# Reson8 open-sources 0.5B speech model for 40+ languages

> Rumble-2-Flash, a half‑billion‑parameter transcription model, is free under Apache 2.0 and runs on any Transformers or vLLM stack.

Oossa · 2026-11-18 · https://oossa.com/en/reson8-open-sources-0-5b-speech-model-for-40-languages

Reson8 announced that its new speech‑to‑text model, Rumble-2-Flash, is now available as a free download with open weights. The model has 0.5 billion parameters and supports automatic detection of more than 40 languages. It can be used through Hugging Face’s 🤗 Transformers library, vLLM, or via Reson8’s own API, which charges €0.01 per hour of audio. The company says the model runs an hour of audio in about half a second, giving it a throughput of over 7,000 real‑time‑factor (RTFx) on full batches.

## How accurate is the tiny model?

On the Hugging Face Open ASR Leaderboard multilingual track, Rumble-2-Flash achieved an average word error rate (WER) of 6.96 % across seven languages, beating Meta’s 7 B‑parameter omniASR‑LLM and OpenAI’s Whisper large‑v3. For English alone the model reported about 5 % WER. In a noisy call‑centre benchmark from AppTek (302 hours, 14 English accents) it scored 7.86 % WER, essentially the same as NVIDIA’s 30 B‑parameter Nemotron 3.5 Transcribe and better than the 0.6 B‑parameter Qwen3‑ASR (8.7 %).

## Why this matters now

Most large speech models are built primarily for English, leaving speakers of other languages with higher error rates. Reson8’s CEO Raoul Ritter notes that languages spoken by millions, such as Dutch or Swedish, are often “an afterthought.” By offering a small, fast, and multilingual model for free, developers can add transcription to apps without paying for huge cloud services or dealing with costly licensing. The low price point (a thousand hours of audio costs €10) also makes it viable for small businesses, educators, and hobbyists who need bulk transcription.

## The facts

- Rumble-2-Flash has 0.5 billion parameters and is released under the Apache 2.0 license.
- The model supports automatic language detection for more than 40 languages.
- It achieves 6.96 % average WER on the Hugging Face Open ASR multilingual benchmark, better than Meta’s 7 B omniASR‑LLM.
- On AppTek’s 302‑hour call‑centre test it scores 7.86 % WER, matching NVIDIA’s 30 B Nemotron 3.5 Transcribe.
- Reson8’s API charges €0.01 per hour of audio, making a thousand hours cost €10.

## Why it matters

For a small company that needs to transcribe customer calls, the model offers near‑enterprise accuracy at a fraction of the cost. For a student or podcaster, it means turning hours of recordings into text in seconds without paying for a large cloud provider. The open‑source release also lets researchers experiment with multilingual speech recognition without building a model from scratch.

## Sources & references

1. [Rumble-2-Flash: an open 0.5B speech model for 40+ languages](https://huggingface.co/blog/rumble-2-flash) – Hugging Face, 2026-11-18
2. [Reson8 releases Rumble-2-Flash speech model as a free open-weights download](https://siliconangle.com/2026/11/18/reson8-releases-rumble-2-flash-speech-model-as-a-free-open-weights-download/) – SiliconANGLE, 2026-11-18
3. [Reson8](https://www.reson8.dev/blog/rumble-2-flash?utm_source=openai) – Reson8

Last updated: 2026-11-18
