Trusence Every claim has a source
Last updated 21 November 2026 Search Türkçe
Models

Reson8 releases Rumble-2-Flash, a 0.5B open speech model covering 40+ languages

Being small and open under Apache 2.0, the model can run on standard Transformers and vLLM setups without custom runtime code.

Reson8 has released Rumble-2-Flash, a 0.5-billion-parameter speech recognition model covering more than 40 languages, under an Apache 2.0 license, with weights on the Hugging Face Hub. Because it is built on the Qwen3-ASR architecture, it runs in standard Transformers and vLLM without custom code, and it offers automatic language detection plus output options such as digit or word numerals, filler words, and forced language. The model was produced by distilling the larger Rumble-2 system into it using on-policy distillation, where the student transcribes audio itself and receives per-token feedback from the teacher. On the Open ASR Leaderboard multilingual track it averages a 6.96% error rate across nine languages, which the sources say beats Meta's 7B omniASR-LLM and Whisper large-v3. On AppTek's 302-hour call-centre benchmark it reaches a 7.86% word error rate, close to NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78%, and on large full batches it runs above 7,000 RTFx.

Why it matters

A compact, openly licensed speech recognition model that matches much larger systems on several benchmarks changes what is available to people who transcribe speech and cannot afford or host the biggest models. Because it loads in standard tooling with no custom code, it can be tried and checked by anyone who follows the field. The benchmark results are reported by the developer and the sources cited, so their weight depends on independent testing of the same datasets.

Sources · 1