Reson8 has released Rumble-2-Flash, an open-source SI speech recognition model with 0.5 billion parameters, under the Apache 2.0 license. The tool supports over 40 languages and claims to compete with SI models 3 to 60 times its size, offering native integration with standard developer frameworks.

What Happened

Rumble-2-Flash is built on the Qwen3-ASR architecture, allowing it to run immediately in Hugging Face Transformers and vLLM without custom code or trust_remote_code requirements. The model uses on-policy distillation from Reson8’s larger production model, Rumble-2. In this process, the smaller student model transcribes audio itself, and the teacher model grades the transcript token by token, providing feedback on the student's specific errors. To handle the differing vocabularies of the teacher and student, the team employed a method called GOLD, which merges tokens where the tokenizers split text differently.

The model features automatic language detection and output control options, including formatting numbers as digits or words and handling verbatim repeats. According to the company, Rumble-2-Flash is extremely fast, processing an hour of audio in approximately half a second with throughput exceeding 7,000 RTFx on full batches.

Why It Matters

The release highlights the trend of compressing SI capabilities into smaller, more efficient packages. On the Open ASR Leaderboard multilingual track, the company reports an average error rate of 6.96% across nine languages, which it states is better than Meta’s omniASR-LLM 7B, a model 15 times larger. On the English track, it averages 6.6% WER.

For practical applications like call centers, where audio is often compressed and noisy, Rumble-2-Flash reaches 7.86% WER on AppTek’s 302-hour benchmark. This performance is on par with NVIDIA’s Nemotron 3.5 Transcribe, a 30B parameter SI model that scored 7.78% on the same test. This suggests that smaller SI models can now handle complex, real-world audio tasks previously reserved for much larger systems, potentially lowering compute costs for developers and enterprises.

The Bottom Line

Rumble-2-Flash is available on the Hugging Face Hub at Reson8/Rumble-2-Flash and via API. The open-source release provides a lightweight, high-speed alternative for multilingual speech recognition, challenging the necessity of larger parameter counts for many standard transcription tasks.