RELEASE · MODELS · #3835
Reson8 open-sources Rumble-2-Flash, a 0.5B multilingual ASR model
Reson8 released Rumble-2-Flash under the Apache 2.0 license: a 0.5-billion-parameter automatic speech recognition model supporting 40+ languages that the team says matches the accuracy of models 3–60x larger. It uses the Qwen3-ASR architecture with a GOLD-enabled on-policy distillation pipeline, runs in Hugging Face Transformers and vLLM without custom runtime changes, achieves over 7,000 RTFx on full batches, posts strong Open ASR Leaderboard and AppTek call-centre results (comparable to NVIDIA Nemotron 3.5 Transcribe on a 302-hour benchmark), and is available on the Hugging Face Hub and via Reson8's API.
KEY POINTS
- Reson8 released Rumble-2-Flash under the Apache 2.0 license: a 0.5-billion-parameter automatic speech recognition model supporting 40+ languages that the team says matches the accuracy of models 3–60x larger.
- It uses the Qwen3-ASR architecture with a GOLD-enabled on-policy distillation pipeline, runs in Hugging Face Transformers and vLLM without custom runtime changes, achieves over 7,000 RTFx on full batches, posts strong Open ASR Leaderboard and AppTek call-centre results (comparable to NVIDIA Nemotron 3.5 Transcribe on a 302-hour benchmark), and is available on the Hugging Face Hub and via Reson8's API.
- A small, fast, permissively licensed ASR model that matches much larger models' accuracy can materially reduce inference cost and broaden deployability of multilingual speech transcription.
WHY IT MATTERS
A small, fast, permissively licensed ASR model that matches much larger models' accuracy can materially reduce inference cost and broaden deployability of multilingual speech transcription.