AI NewsModels & agentsAnnouncement

Reson8 released Rumble-2-Flash, a 0.5B open speech model for 40 languages

Reson8 released Rumble-2-Flash, a 0.5B open speech recognition model under Apache 2.0, with the Qwen3-ASR architecture and native Transformers support.

AI News

Editorial2 min read

LinkedInX
Reson8 diagram of the GOLD on-policy distillation method used for Rumble-2-Flash

Image: Reson8

Why it mattersA speech model that fits on a small server reads 40 languages. That opens a path off paid speech APIs for self-hosted teams.

A speech model that fits on a single accelerator is new ground. It lets teams leave a hosted API. Reson8 published Rumble-2-Flash on Hugging Face on 18 November. The authors are Mark 't Hart, Eustache Le Bihan and Eric Bezzam. The model has 0.5B parameters. It is released under Apache 2.0. The weights are at Reson8/Rumble-2-Flash.

Reson8 built the model on the Qwen3-ASR architecture. That choice gives it native support in Hugging Face Transformers and vLLM. A team does not have to write custom inference code to use it. The post calls out automatic language detection across 40 or more languages.

The benchmark numbers Reson8 published

On the multilingual Open ASR Leaderboard, Rumble-2-Flash averages 6.96 percent word error rate across nine languages. Those are Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. Reson8 says this is "better than Meta's omniASR-LLM 7B, a model 15x larger, and ahead of Whisper large-v3". The English-only track is 6.6 percent.

On AppTek's call-centre benchmark, 302 hours of helpline calls in 14 English accents, Rumble-2-Flash scores 7.86 percent. NVIDIA's Nemotron 3.5 Transcribe, a 30B model, scores 7.78 percent on the same test. Qwen3-ASR scores 8.7 percent at 0.6B and 7.4 percent at 1.7B. On full batches Reson8 reports more than 7,000 RTFx, which is one hour of audio in about half a second.

What GOLD changes

Reson8 trained the small model with a method it calls GOLD. The problem it solves is that on-policy distillation usually needs the teacher and the student to share a tokenizer. Reson8 picked Qwen3-ASR for the student, so it also inherits that tokenizer. The teacher used different tokens. GOLD merges tokens where the two split the text differently. Both models can then be scored over the same text.

The training itself is on-policy. The student does not train on transcripts the teacher wrote. The student transcribes the audio itself. The teacher then grades every token the student produced and gives its own probability for the next token. The student learns to match the teacher position by position.

What a team can control

The model takes a transcription request that can force the output format. A team can ask for numbers as digits or as words. It can keep fillers in or drop them. It can force the language rather than detect it. The API is available through Reson8 directly for a team that would rather not host the weights.

Source

Rumble-2-Flash: Unlocking On-Policy Distillation for Any Model Family, Reson8.

SourceReson8

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project