Models·
Rumble-2-Flash: A 0.5B Open Speech Model That Rivals Giants
Rumble-2-Flash is a 0.5B parameter open-source speech recognition model for 40+ languages, delivering near-large-model accuracy with extreme inference speed on standard GPU stacks.

Hugging Face and Reson8 have released Rumble-2-Flash, a 0.5-billion-parameter automatic speech recognition model covering more than 40 languages under the permissive Apache 2.0 license.
The model integrates natively with Hugging Face Transformers and vLLM, requiring no custom runtime code. It offers automatic language detection, configurable output formatting (digits, fillers, verbatim repeats), and delivers inference speeds exceeding 7,000 real-time factor on full batches — transcribing an hour of audio in roughly half a second.
Why it matters for GPU infrastructure
Rumble-2-Flash achieves its performance through on-policy distillation with the GOLD algorithm, transferring knowledge from a larger teacher model token by token. This approach allows the compact student to match the accuracy of models 15 to 60 times its size, with a multilingual error rate of 6.96% and 6.6% word error rate on English benchmarks.
For AI infrastructure teams, the small footprint translates directly into lower GPU memory requirements and higher throughput per dollar. The model can run real-time transcription on modest accelerators, enabling cost-effective deployment in call centers, edge devices, and multilingual applications without sacrificing quality.
- aigpu
- ai gpu
- ai gpu cloud
- aigpu dubai
- speech recognition
- open source
- small model
- inference
By AiGpu Editorial · Editorial rewrite based on public reporting (Hugging Face Blog)
← All articles