Skip to main content
Back

Reson8 Opens Rumble-2-Flash, a 1GB Speech Model Supporting 45 Languages

Reson8's 0.5B open-weights speech-to-text model lands on Hugging Face under Apache 2.0, beating Whisper Large V3 at a third of its size across 45 languages.

Model, Full Access
  • Rumble-2-Flash is a 0.5B open-weights ASR model under Apache 2.0, covering 45 languages.
  • Averages 7.09% WER multilingual and ~5% English, beating Whisper Large V3 on low-resource languages.
  • Distilled on-policy from the larger Rumble-2 teacher, using Qwen3-ASR architecture as the base.
  • Runs natively on Transformers, vLLM, llama.cpp (GGUF), and Apple Neural Engine via Core ML.
  • Tuned for prerecorded audio with accents, phone calls, overlapping speech and background noise.
  • Weights are ~1 GB, small enough to run on a phone or laptop without a GPU.

Amsterdam-based Reson8 has released Rumble-2-Flash, a 500 million-parameter speech recognition model, as open weights under the Apache 2.0 license. The model card lists 45 languages, and its weights occupy roughly 1 GB at 16-bit precision. Reson8 reports competitive accuracy against several larger multilingual systems, giving developers a smaller option for self-hosted transcription on servers, laptops, and converted mobile runtimes.

Reson8’s release post identifies the company as Resonate Labs B.V. It previously offered Flash through a paid API, beginning on Nov. 3, and launched its larger Rumble-2 model one week later. The open release permits commercial use, modification, and redistribution under Apache 2.0’s license conditions.

Benchmark gains, with caveats

Reson8 reports a 6.60% average English word error rate on the Open ASR Leaderboard. Word error rate, or WER, measures substitutions, deletions, and insertions relative to a reference transcript, so lower scores indicate fewer transcription errors.

Evaluation Rumble-2-Flash WER Comparison or context
Open ASR English average 6.60% Reported from the public leaderboard suite
Seven-language multilingual run 7.09% Includes German, French, and Dutch; English scored about 5%
Ten lower-resource languages 17.6% Whisper Large V3 scored 39.82% in Reson8’s evaluation, which included Maltese and Basque