
Reson8 Releases Rumble-2-Flash Speech Recognition Weights Under Apache 2.0
The AMW Read
The release adds a self-hosted option for multilingual transcription, with a developer-reported benchmark and a deliberate open-weight distribution choice.
Reson8 Releases Rumble-2-Flash Speech Recognition Weights Under Apache 2.0
Reson8 has released the weights of Rumble-2-Flash, a speech recognition model designed to transcribe recorded audio, under the Apache 2.0 license. The model has about 500 million parameters and supports more than 40 languages, according to the report. It can identify the spoken language and lets developers choose how to render numbers and whether to retain hesitations and repetitions. Reson8 had previously offered the model through an API.
The release gives developers another way to run speech recognition in their own environments, using existing Transformers or vLLM infrastructure. Rumble-2-Flash uses the Qwen3-ASR architecture and was trained with Reson8's larger Rumble-2 model as a teacher. In an evaluation published by the development team, it recorded a 7.86% word error rate across 302 hours of call-center audio covering 14 English accents. A roughly 30-billion-parameter NVIDIA model scored 7.78% on the same reported evaluation.
For teams considering subtitles or call records, the comparison is useful but does not establish accuracy across every language or recording condition. Developers should test the model on their own speakers, audio quality, and target languages. They should also measure individual recordings separately from batch workloads: the reported high processing speed depended on grouping enough recordings together. Self-hosting also changes the deployment choice for customers who previously accessed the model through Reson8's API.