Clean audio is easy. A phone call is the other exam: compressed, noisy, with speakers who have very different accents. That is exactly where Reson8 measures its new model, and it says Rumble-2-Flash, with 0.5 billion parameters, keeps pace with NVIDIA's 30B Nemotron.
Fact
The facts: on 18 November 2026 Reson8 announced on the Hugging Face blog that it is open sourcing Rumble-2-Flash under the Apache 2.0 license. It is a speech recognition model with 0.5 billion parameters for 40+ languages that, according to Reson8, competes with models 3-60x its size. It uses the Qwen3-ASR architecture and runs in Transformers and vLLM with no custom code. On the multilingual track of the Open ASR Leaderboard, according to Reson8, the average error rate is 6.96% across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English: better than Meta's omniASR-LLM 7B, a model 15x larger, and ahead of Whisper large-v3. On the English track the average is 6.6% WER. On speed, Reson8 reports over 7,000 RTFx on full batches, which is an hour of audio in about half a second. On AppTek's benchmark (302 hours of helpline calls, 14 English accents) Rumble-2-Flash reaches 7.86% WER, on par with Nemotron 3.5 Transcribe (30B) at 7.78%. Qwen3-ASR 0.6B scores 8.7% and Qwen3-ASR 1.7B scores 7.4%. The weights are on the Hub as Reson8/Rumble-2-Flash, and the model is also available through the company's API.
On par is not better. Nemotron is at 7.78%, Rumble-2-Flash at 7.86%: slightly behind, at 0.5 billion parameters against 30. The leaderboard and the test are external, but the results are reported by Reson8 itself. The speed has a condition too: half a second for an hour of audio holds for full batches, and throughput drops when they are less full.
How it was built is the more interesting part. Reson8 took its larger production model, Rumble-2, and distilled it into the small one. The student does not learn from transcripts the teacher wrote. It transcribes the audio itself, and the teacher grades the transcript token by token: at every position it gives its own probability for what comes next, and the student learns to match it. So it gets corrected on exactly the mistakes it makes. There was an obstacle: the two models do not share a tokenizer, because Qwen3-ASR brings its own vocabulary. A method called GOLD merges tokens where the two tokenizers split the text differently, so both models are compared over the same text.
The student writes on its own. The teacher grades every token.
And the other languages? The post does not give the list of languages: only the nine from the multilingual test are named. Which ones are among the rest, the source does not say, so neither do we. It does not give API pricing either, or the size of Rumble-2.
Download the weights and run your own recording in your own language through the model, from a phone, not a studio. Until the language list appears, there is no other check.
Official primary sources
Reson8 - Rumble-2-Flash: an open 0.5B speech model for 40+ languages, 18.11.2026The visual is generated code-art, no third-party images.