
Rumble-2-Flash: A Tiny Open Speech Model for 40+ Languages
Rumble-2-Flash is a 0.5B open speech recognition model for 40+ languages. Reson8 says it can transcribe an hour of audio in about half a second.
Speech recognition usually forces a trade between size and accuracy. Rumble-2-Flash, released by Reson8 under the Apache 2.0 license, is an attempt to loosen that trade. The open speech recognition model has just 0.5 billion parameters, supports more than 40 languages with automatic language detection, and runs through standard Hugging Face tooling. For developers building transcription into call analytics, meeting notes or voice agents, that is a compact and very approachable combination.
- Size and license: 0.5B parameters, open weights, Apache 2.0.
- Languages: 40+ with automatic detection and an option to force a language.
- Speed: over 7,000 times real time on full batches, or about half a second per hour of audio, according to Reson8.
- Phone-call accuracy: 7.86% word error rate on AppTek's 302-hour call-center benchmark, as reported by Reson8.
What Is Rumble-2-Flash?
Rumble-2-Flash is a distilled version of Rumble-2, Reson8's larger production model. It uses the Qwen3-ASR architecture, which gives it native day-one support in Transformers and vLLM without a forked runtime or custom remote code. That detail may be the most useful one for engineering teams, because a model is far easier to evaluate when it drops into an existing serving stack.
The output controls are practical too. Developers can choose digits or spelled-out numbers, keep or remove filler words, preserve verbatim repeats and force a target language when automatic detection is not wanted.
How Does On-Policy Distillation Work?
The training method is the technically interesting part. In conventional distillation, a small student model often learns from transcripts produced by a larger teacher. Reson8 instead has the student transcribe the audio itself, then asks the teacher to grade that attempt token by token. Because the feedback targets the student's own mistakes, each position in the transcript becomes a learning signal.
There was one complication: the two models do not share a tokenizer. Reson8 used a technique called GOLD, which merges tokens where the vocabularies split text differently, so teacher and student can still be compared over the same words.
How Accurate Is Rumble-2-Flash?
Reson8 reports a 6.96% average error rate on the Open ASR Leaderboard's multilingual track, covering Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, and 6.6% on the English track. The company says that places it ahead of Whisper large-v3 and Meta's much larger omniASR-LLM 7B on the multilingual track.
The call-center result is a useful stress test because phone audio is compressed, noisy and full of varied accents. On AppTek's benchmark of helpline calls across 14 English accents, Reson8 reports 7.86% for Rumble-2-Flash, close to 7.78% for NVIDIA's far larger Nemotron 3.5 Transcribe. These are vendor-reported comparisons, so it is worth running your own audio before committing. Domain terms, microphones and accents can change the picture considerably.
Who Should Try This Open Speech Model?
Teams processing large audio archives have the clearest reason to test it, since throughput directly affects cost. Developers of multilingual voice products should also take a look, especially if they want weights they can host themselves. It joins a lively field that includes Salesforce's open SOV Transcribe model, and Hugging Face's growing inference options make side-by-side trials easier. You can follow more releases like this in our AI news section.
Sources: Reson8 on the Hugging Face Blog — November 18, 2026; Rumble-2-Flash model page — November 18, 2026.
More AI Stories

Prompt Caching Explained: How to Cut LLM API Input Costs
Prompt caching lets LLM APIs reuse a repeated prefix at up to 95% off. Learn how Claude, OpenAI, Gemini and Qwen handle it and how to structure prompts.

GPT-6.2 Astra and Sol: Prices, Scores and Who Should Switch
GPT-6.2 Astra drops to $8 per million input tokens as Sol holds at $2. Here is what changed, what independent tests show and which model fits your work.

Gemini Google Wallet Integration: AI Budgeting Explained
Gemini can now read your Google Wallet spending to chart budgets, flag pricier subscriptions and plan holiday costs. Here's how the US rollout works.
