Reson8 Open-Sources Rumble-2-Flash, an Apache 2.0 0.5B Speech Model for 40+ Languages

Amsterdam-based speech startup Resonate Labs B.V.—better known as Reson8—released its 500-million-parameter speech recognition model, Rumble-2-Flash, as an open-weights download under the Apache 2.0 license on November 18, 2026. Originally launched in Reson8's hosted API on November 3 at €0.01 per hour of audio, the 0.5B model supports more than 40 languages with automatic language detection and ships with native day-0 support in Hugging Face Transformers and vLLM. For developers running local AI pipelines on Linux workstations or resource-constrained servers, the release shows how distillation across model families is shrinking multilingual speech-to-text down to sub-billion-parameter footprints. ### Cracking the Small-Model Accuracy Ceiling With GOLD Small automatic speech recognition (ASR) models have traditionally traded accuracy for latency, stumbling badly once audio leaves quiet studio recordings for compressed telephone lines or non-English accents. To prevent that drop-off without keeping Rumble-2-Flash locked to a bespoke runtime, Reson8 built the 0.5B model on Alibaba's Qwen3-ASR architecture—meaning any stack already running Transformers or vLLM can load `Reson8/Rumble-2-Flash` directly without forked code or `trust_remote_code` flags. Bridging Reson8's larger proprietary teacher model (`Rumble-2`) to a Qwen3-ASR student created a technical hurdle: standard on-policy distillation normally requires teacher and student models to share the exact same tokenizer. According to the engineering post published on Hugging Face by Reson8 researchers alongside co-authors Eustache Le Bihan and Eric Bezzam, the team solved that mismatch using GOLD—a token-merging method that aligns disparate tokenizers over the same underlying text so the teacher can grade the student's own generated transcripts token by token. Rather than training passively on static ground-truth text, the student learns directly from probability feedback on the exact errors it makes during decoding. ### Benchmarks, Throughput, and What Remains Unverified On Hugging Face's Open ASR Leaderboard multilingual track—covering Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English—Rumble-2-Flash averages a 6.96% word error rate (WER) and 6.6% WER on the English track, outperforming Meta's 7B `omniASR-LLM` and OpenAI's Whisper Large v3 while running at over 7,000x real-time factor (RTFx) in full-batch evaluation, roughly an hour of audio in half a second. In a separate evaluation across 10 lower-resource languages including Basque and Maltese cited by SiliconANGLE, Reson8 reported a 17.6% mean WER compared with 39.82% for Whisper Large v3. Telephony audio provides a sterner test of real-world robustness. On AppTek's 302-hour call-center benchmark spanning 14 English accents—scored using AppTek's official evaluation script—Rumble-2-Flash posted a 7.86% WER. That places the 0.5B open model nearly level with NVIDIA's 30B Nemotron 3.5 Transcribe (7.78% WER), ahead of Qwen3-ASR 0.6B (8.7%), Mistral's Voxtral Small (9.12%), and IBM's Granite Speech 5.0 2B (11.07%), trailing only Qwen3-ASR 1.7B at 7.4%. At the same time, most of these comparative figures come from Reson8's own benchmark runs rather than third-party audits, and independent community verification on noisy non-English telephony sets will matter before enterprises swap out established pipelines. ### Open-Core Economics in European Voice AI Releasing a competitive 0.5B speech model under a permissive Apache 2.0 license raises an obvious commercial question: why give away a model that undercuts API rivals? Reson8's product split reveals a deliberate open-core boundary. Rumble-2-Flash targets batch transcription of prerecorded files—podcasts, voice notes, call archives, and lectures—offering prompt-level control tokens for digits versus spelled-out numbers, verbal fillers, and verbatim repeats. By contrast, real-time streaming, speaker diarization, and Reson8's flagship ability to inject up to 50,000 domain-specific medical, legal, or financial terms without fine-tuning remain exclusive to the commercial `Rumble-2` model and its EU-hosted, zero-retention infrastructure. By seeding developers with a fast, Apache 2.0 baseline that runs locally out of the box, Reson8 aims to pull European and multilingual builders into its ecosystem first—and monetize regulated, live enterprise workloads later. ### Sources & Further Reading - [Hugging Face Blog (Technical release and GOLD distillation details)](https://huggingface.co/blog/rumble-2-flash) - [Reson8 Official Blog (Company product benchmarks and open-core positioning)](https://www.reson8.dev/blog/rumble-2-flash) - [SiliconANGLE (US enterprise AI coverage and low-resource language figures)](https://siliconangle.com/2026/11/18/reson8-releases-rumble-2-flash-speech-model-as-a-free-open-weights-download/) - [TokenPost (International technology coverage)](https://www.tokenpost.com/news/technology/60508) ~AI ~OpenWeights ~SpeechRecognition ~LocalAI

Reson8 Open-Sources Rumble-2-Flash, an Apache 2.0 0.5B Speech Model for 40+ Languages
0
Discuss 0 3
0 replies
Be the first to join this thread.