Reson8 Distills Its Production Speech Model Into Rumble-2-Flash, a 0.5B Apache-2.0 Recognizer That Holds Up on Phone Calls
European voice-AI company Reson8 released Rumble-2-Flash on November 18, 2026, a 0.5B-parameter speech recognition model for 40-plus languages under Apache-2.0. It runs natively in Transformers and vLLM, averages 6.96% error on the Open ASR Leaderboard's multilingual track and, by Reson8's measure, matches a 30B NVIDIA model on a 302-hour call-centre benchmark.
Reson8 has open-sourced a compact version of its production speech recognition model. On November 18, 2026, the European voice-AI company released Rumble-2-Flash, a 0.5B-parameter model that transcribes more than 40 languages with automatic language detection, under the Apache 2.0 license. The weights are on Hugging Face at Reson8/Rumble-2-Flash, and the model is also available through Reson8’s API.
Reson8 says the model “competes with models 3-60x its size.” Its model card lists 45 language codes, from Afrikaans and Arabic to Vietnamese and Cantonese.
No custom runtime required
The release leans heavily on compatibility. Rumble-2-Flash uses the Qwen3-ASR architecture, which Reson8 says gives it day-zero support in Hugging Face Transformers and vLLM without a forked runtime or trust_remote_code. The model card shows a Transformers quickstart and a vLLM setup that exposes the model through an OpenAI-compatible chat endpoint, with audio passed in as base64.
Four transcription switches are exposed through the chat template. Users can force a language by name, write numbers as digits (the default) or words, keep fillers such as “um” and “uh” (off by default), and choose whether to keep verbatim repeats (on by default) or collapse them. If no language is set, output begins with the detected language name.
Distilled from the company’s bigger model
Rumble-2-Flash is a compressed version of Rumble-2, Reson8’s production model, which the company launched in its API on November 10 across more than 50 languages. To find out how small a model could get before accuracy fell off, Reson8 used on-policy distillation. Instead of training the small model on transcripts written by the large one, the student transcribes audio itself and the teacher grades each token by giving its own probability for what should come next. Reson8 says this corrects the student on exactly the mistakes it makes, with feedback on every token rather than one score per clip.
That method normally requires teacher and student to share a tokenizer, and these two do not, because adopting Qwen3-ASR meant adopting its vocabulary. Reson8 says it used GOLD, a technique that merges tokens where the two tokenizers split text differently so both models are compared over the same text, matching shared tokens directly and the rest by sorted probability.
Benchmarks on clean and noisy audio
On the Hugging Face Open ASR Leaderboard, evaluated on the November 18 release (v1.16.1) on an H200 with vLLM, Rumble-2-Flash averages a 6.96% error rate on the multilingual track across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. Reson8 says that beats Meta’s omniASR-LLM 7B, a model 15 times larger, and Whisper large-v3. On the English track it averages 6.6% word error rate.
The per-language spread is wide. Spanish averages 3.73% and Italian 5.09%, while French reaches 7.21% and Hindi 14.11%. By Reson8’s own figures, the full-size Rumble-2 scored 8.95% on Hindi and 2.47% on Spanish.
Speed is the trade in the other direction. On the leaderboard’s larger datasets, where batches stay full, Reson8 says the model exceeds 7,000 RTFx, or about an hour of audio in half a second. The English-track average is 6,338 RTFx.
Reson8 argues that phone audio is the harder test because it is compressed, noisy and full of varied accents. On AppTek’s call-centre benchmark, 302 hours of helpline calls across 14 English accents scored with AppTek’s own script, Rumble-2-Flash reaches 7.86% WER. Reson8 puts that on par with NVIDIA’s 30B Nemotron 3.5 Transcribe at 7.78%, ahead of Qwen3-ASR 0.6B at 8.7%, and behind only Qwen3-ASR 1.7B at 7.4% among the models it compared. The call-centre comparison was run and published by Reson8.
EU-hosted GPUs and a Rumble-2 API from €0.14 an hour
Reson8 sells speech-to-text to regulated industries such as healthcare, legal and finance. It says it processes audio only on GPUs in the EU, specifically its own in the Netherlands, with zero audio retention and no customer data used for training. Its API offers prerecorded, realtime and turn-aware endpoints, the last with barge-in detection for voice agents, and its pricing page lists Rumble-2 from €0.14 an hour.
The larger Rumble-2 remains an API product. The model card says nothing about streaming support for the open model; the examples shown transcribe complete audio files. Reson8 has asked users to share results and requests in the model’s discussion tab.
- #Reson8
- #Rumble-2-Flash
- #Speech Recognition
- #Open Source
- #Hugging Face
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.