Reson8 open-sourced Rumble-2-Flash on 18 November 2026 under the Apache 2.0 licence. It is a 0.5-billion-parameter automatic speech recognition (ASR) model covering 40+ languages, built on the Qwen3-ASR architecture, with Hugging Face and vLLM support. The model is 1.39 GB.
One date needs care: 18 November is the open-weights release. The model itself debuted in Reson8's API earlier, on 3 November 2026 — so 18 November is the day the weights went public, not the day the model first appeared.
Reson8's own numbers, from its Open ASR Leaderboard, are a 6.96% average error rate across languages, 6.6% word error rate (WER) on English, and 7.86% WER on the AppTek 302-hour call-centre benchmark, with throughput above 7,000 RTFx. These are Reson8's own runs, not independent measurements. The hosted API starts from €0.01 per hour of audio.
What it does not do matters too: there is no speaker diarization, no end-of-turn detection and no automatic language detection — those features are reserved for the larger Rumble-2 model.
Key facts
| Item | Details |
|---|---|
| Model | Rumble-2-Flash, 0.5B-parameter ASR model |
| Open-weights release | 18 November 2026, Apache 2.0 |
| API debut | 3 November 2026 (earlier than the open-weights release) |
| Languages | 40+ |
| Architecture | Qwen3-ASR |
| Support | Hugging Face + vLLM |
| Size | 1.39 GB |
| Reson8's leaderboard figures | 6.96% avg error (multilingual); 6.6% WER (English); 7.86% WER (AppTek 302-hour call-centre benchmark); >7,000 RTFx |
| API pricing | From €0.01 per hour of audio |
| Not included | Speaker diarization, end-of-turn detection, automatic language detection (reserved for larger Rumble-2) |
What this means
A 0.5B-parameter speech model that fits in 1.39 GB is aimed at a practical niche: on-device or low-cost transcription in many languages, without renting a large hosted model. The Apache 2.0 licence means developers can use, modify and ship it with the usual open-source terms.
The limits are worth knowing before adopting it. If your use case needs to know who is speaking, when a speaker has finished, or which language is being spoken, Rumble-2-Flash does not do that — Reson8 keeps those features for the larger Rumble-2. And the accuracy figures are Reson8's own leaderboard runs, so treat them as the vendor's numbers until independent tests appear.
A note on verification
The release date, licence, architecture and benchmark figures are as reported for the 18 November 2026 open-weights release; the 3 November API debut is recorded separately so the two dates are not confused. The leaderboard figures are Reson8's own runs.
FAQs
When was Rumble-2-Flash released?
The open-weights release was 18 November 2026 under Apache 2.0. The model had already debuted in Reson8's API on 3 November 2026.
How many languages does it cover?
40+ languages, on the Qwen3-ASR architecture, with Hugging Face and vLLM support.
What does it not do?
No speaker diarization, no end-of-turn detection and no automatic language detection — those are reserved for the larger Rumble-2 model.
Are the accuracy figures independently verified?
No. The 6.96% average error, 6.6% English WER and 7.86% AppTek WER are Reson8's own leaderboard runs.
Source: Reson8 open-weights release, 18 November 2026; Reson8 Open ASR Leaderboard. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!