Reson8 open-sources Rumble-2-Flash, a 0.5B-parameter multilingual speech recognition model under Apache 2.0
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Reson8 released the weights of Rumble-2-Flash (0.5B parameters, Apache 2.0) on the Hugging Face Hub at Reson8/Rumble-2-Flash, and also offers it through its API. It supports 40+ languages with automatic language detection and output controls. It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM without custom code. It was distilled from the larger Rumble-2 model using on-policy distillation with GOLD, which handles mismatched tokenizers. Self-reported results: 6.96% average WER on the Open ASR Leaderboard multilingual track (ahead of omniASR-LLM 7B and Whisper large-v3), 6.6% WER on the English track, over 7,000 RTFx on full batches, and 7.86% WER on AppTek's 302-hour call-centre benchmark (Nemotron 3.5 Transcribe 30B scores 7.78%).
Entities: Reson8, Rumble-2-Flash, Rumble-2, Hugging Face, Qwen3-ASR, NVIDIA Nemotron 3.5 Transcribe
1 primary
What happened
Reson8 released the weights of Rumble-2-Flash, a 0.5B-parameter speech recognition model covering 40+ languages, under Apache 2.0 on the Hugging Face Hub. It is also sold through Reson8's API. It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM without custom code, and was distilled from the larger Rumble-2 model. Reson8 reports 6.96% average word error rate (WER) on the Open ASR Leaderboard multilingual track, 6.6% on the English track, over 7,000 RTFx (speed relative to real time) on full batches, and 7.86% on AppTek's 302-hour call-centre set. All of these figures are self-reported.
Why it matters
Developers who need multilingual transcription on modest hardware get a permissively licensed option that is simple to deploy, so it is worth a quick trial against Whisper large-v3 on their own audio. Call-centre and telephony teams have the most to gain, because the claimed 7.86% on noisy calls is close to a much larger NVIDIA model's 7.78%. The wider impact is limited: Reson8 is a small lab, speech recognition is crowded, and nothing here has yet been independently reproduced.
What is noise
The "competes with models 3-60x its size" line is packaging: it rests on self-reported averages, and a 0.1 point gap against a 30B model on one call-centre set is a tie, not a win. The speed figure is a best case with full batches, so it says little about single-stream or low-latency use. Averages across 40+ languages can also hide weak performance in individual languages, and the post gives no per-language breakdown beyond the headline.
Watch next
- 01Independent runs or leaderboard entries that reproduce the 6.96% multilingual and 7.86% call-centre WER figures, ideally with per-language results.
- 02Real-world latency and accuracy reports on streaming, accented, noisy and low-resource-language audio, compared with Whisper large-v3 and Qwen3-ASR.
- 03Adoption signals over the next 4 to 8 weeks: Hugging Face downloads, community fine-tunes, inclusion in vLLM and other inference tooling, and whether the larger Rumble-2 is opened up too.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680