What changed
Reson8 released Apache-2.0 weights for Rumble-2-Flash, a 0.5B speech-recognition model covering more than 40 languages with native Transformers and vLLM support and configurable transcription output.
Why it matters
A small permissively licensed model can make multilingual transcription practical on constrained or private infrastructure, but reported error rates and throughput need matched audio, hardware and decoding settings.
What to do
Pin the model and runtime revisions, transcribe a consented multilingual holdout with noisy calls and domain terms, and compare word error rate, latency, memory and failure handling against the current baseline.
Caveat
Leaderboard, throughput and call-centre results are author-reported and include a project benchmark script. Language coverage does not guarantee equal accuracy, diarization, privacy or robustness in each deployment.