Rumble-2-Flash: an open 0.5B speech model for 40+ languages
- Source
- huggingface.co
- Date

Context
Reson8 has open-sourced Rumble-2-Flash, a 0.5-billion-parameter speech recognition model for more than 40 languages, and its announcement reports accuracy close to models many times its size under the Apache 2.0 licence. It reuses the Qwen3-ASR architecture from Qwen, so it runs in the standard Transformers and vLLM libraries without custom code, detects the spoken language automatically, and can be told whether to write numbers as digits or words and whether to keep fillers such as um.
Reson8 got a model this small to stay accurate through on-policy from Rumble-2, its larger production model. Instead of training on transcripts the larger teacher model had written, the small student transcribes the audio itself, and the teacher scores that attempt word piece by word piece, giving its own probability for what should come next at every position. The student learns to match those probabilities, so it is corrected on exactly the mistakes it makes. That method normally requires both models to split text into the same tokens, which these two do not, so Reson8 used GOLD, a technique that aligns the two vocabularies by merging tokens where they split text differently.
These results are as reported in the announcement. On the multilingual track of Hugging Face's Open ASR Leaderboard, the model averages a 6.96 percent word error rate, the share of words transcribed wrongly, across nine languages including Chinese, Hindi and English, which Reson8 says beats Meta's 7-billion-parameter omniASR- and OpenAI's Whisper large-v3. On AppTek's call-centre benchmark of 302 hours of compressed, noisy helpline calls in 14 English accents, it scored 7.86 percent against 7.78 percent for NVIDIA's Nemotron 3.5 Transcribe, a result Reson8 calls on par, and it names Qwen3-ASR 1.7B, at 7.4 percent, as the only model in its comparison that does better. The quoted speed of over 7,000 times real time, about an hour of audio in half a second, applies only when batches are full.
The size comparisons need one adjustment. Reson8 lists Nemotron 3.5 Transcribe at 30 billion parameters, sixty times Rumble-2-Flash's size, but that model is a design that uses only about 4 billion parameters per step, according to Artificial Analysis's evaluation, so the computation it does per step is roughly eight times larger, not sixty. Even so, a half-billion-parameter model matching a far larger one on noisy phone audio is a meaningful result if independent runs confirm it, and the and standard architecture let anyone test it on their own recordings.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Checking sign-in…
Loading comments…


