Open AI Models

Reson8 open-sources Rumble-2-Flash, 0.5B-parameter speech model rivaling 30B rivals

Key Takeaway

A 0.5B-parameter Apache-2.0 speech recognition model achieves near-parity with NVIDIA's 30B Nemotron 3.5 Transcribe on call-center audio through on-policy distillation with cross-tokenizer GOLD alignment.

Summary

Reson8 has released Rumble-2-Flash, a 0.5-billion parameter automatic speech recognition model available under the Apache 2.0 license with native support in Hugging Face Transformers and vLLM. The model handles 40+ languages with automatic detection, and offers configurable output formatting including digit versus word numbers, filler word handling, and verbatim repeat preservation. On the Open ASR Leaderboard multilingual track, it achieves a 6.96% word error rate across nine languages, outperforming Meta's 7B omniASR-LLM and Whisper large-v3 despite being 15 times smaller.

The model's architecture uses Qwen3-ASR for immediate ecosystem compatibility without custom runtimes or trust_remote_code requirements. Its training employed on-policy distillation from Reson8's larger production model, where the student generates transcripts and receives per-token probability feedback from the teacher rather than training on static teacher outputs. A technique called GOLD enabled this cross-tokenizer distillation by aligning differently-tokenized outputs for direct comparison. Inference throughput exceeds 7,000 real-time factors on full batches, processing one hour of audio in approximately half a second.

Architecture and distillation approach

  • Qwen3-ASR backbone: Rumble-2-Flash adopts the Qwen3-ASR architecture, enabling day-zero deployment through standard Hugging Face Transformers and vLLM pipelines without forked runtimes or trust_remote_code execution.
  • On-policy distillation with GOLD: The student model transcribes audio independently and receives per-token probability distributions from the teacher as training signal, learning from its own mistakes rather than static teacher transcripts; GOLD merges divergent tokenizer outputs to enable cross-vocabulary comparison without shared tokenizers.
  • Throughput scaling: Batch fullness directly determines speed, with full batches achieving over 7,000 RTFx on Open ASR Leaderboard datasets, translating to approximately 0.5 seconds for one hour of audio.
  • Call-center robustness: On AppTek's 302-hour benchmark with 14 English accents and compressed noisy audio, the model reaches 7.86% WER, statistically matching NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78% and exceeding Qwen3-ASR 0.6B's 8.7%.

Why it matters

Rumble-2-Flash demonstrates that aggressive distillation can collapse the parameter gap between deployable edge models and datacenter-scale speech systems, potentially redirecting developer investment toward efficient open-weight models rather than proprietary APIs. Its Apache-2.0 licensing and standard Transformers integration remove deployment friction for Linux-based inference pipelines, while the GOLD technique for cross-tokenizer distillation may generalize to other model compression scenarios in the open-source AI stack.

What to do

  • Install torch, transformers, and accelerate using 'pip install torch transformers accelerate'.
  • Use the 'Reson8/Rumble-2-Flash' model ID with Hugging Face Transformers or vLLM for transcription.

Original reference: huggingface.co →