Rumble-2-Flash Is a 0.5B Speech Model That Punches Way Above Its Weight
Reson8 open-sourced Rumble-2-Flash, a 0.5-billion-parameter speech recognition model for more than 40 languages that beats models 60 times its size on public benchmarks.

A tiny open speech model with outsized results
Reson8, a ten-person speech startup from Amsterdam, open-sourced Rumble-2-Flash on November 18, and the numbers are hard to ignore. The model has 0.5 billion parameters, covers more than 40 languages, and the company says it competes with models 3 to 60 times its size. The weights are free to download on Hugging Face under an Apache 2.0 license, which means anyone can run it, modify it, or ship it in a commercial product without paying Reson8 a cent.
The model is an automatic speech recognition system, the kind that turns recorded audio into text. It is built on the Qwen3-ASR architecture and was trained by distilling Reson8's larger Rumble-2 model down into a small one, using a technique the company calls GOLD to line up the tokenizers between teacher and student. The whole package is 1.39 GB, small enough to run on a laptop, a phone, or an on-prem server, and it works with Hugging Face Transformers and vLLM out of the box, with no custom code required.
The benchmark numbers behind the claim
Reson8's headline result comes from the Hugging Face Open ASR Leaderboard, where Rumble-2-Flash averaged a 6.96 percent error rate across nine languages: Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English. That puts it ahead of Meta's omniASR-LLM 7B, a model with roughly 15 times as many parameters, and ahead of OpenAI's Whisper large-v3, the long-standing default for open transcription.
A second benchmark tells a similar story on harder audio. On AppTek's test of 302 hours of hotline calls across 14 English accents, Rumble-2-Flash scored a 7.86 percent error rate, essentially tied with NVIDIA's Nemotron 3.5 Transcribe at 30 billion parameters, which scored 7.78 percent. The best row in that table, Qwen3-ASR 1.7B at 7.4 percent, is a bigger model too. The point is not that a 0.5B model wins every test. It is that it is close enough to models 60 times its size that the size gap stops mattering for most real work.
Benchmarks from a vendor always deserve a skeptical read, and Reson8's numbers are no exception. The company has not published details of its training data, so independent researchers cannot yet check what the model was trained on. But the Open ASR Leaderboard is public, the weights are public, and anyone can download the model and run their own audio through it. That is the difference between a marketing claim and a testable one.
Why a small model matters more than a big one
The speech recognition market has a quiet problem: most of the good models are either huge, English-first, or locked behind an API. A team building transcription for customer support calls in Dutch, or a hospital that cannot send patient audio to a cloud service, or a developer in a country where the big APIs are slow and expensive, has had few good options. Rumble-2-Flash is aimed squarely at that gap.
Running your own model changes the economics. Reson8's hosted API starts at 0.01 euros per hour of audio, which is already cheap, but the open weights cost nothing per hour beyond the hardware you already own. For a call center transcribing thousands of hours a month, or a product that needs transcription on the device with no network at all, that difference is the whole business case. The Apache 2.0 license matters here too. It is permissive, it allows commercial use, and it means a startup can build on the model without a lawyer in the room.
The release also fits a pattern that has been building all year. Small, specialized models keep beating the assumption that bigger is always better. Kimi K3.1 Pro showed open weights can compete at the top end, and Gemini 4 Neon showed Google itself is now shipping models tuned for speed and cost rather than raw capability. Rumble-2-Flash is the same idea applied to speech: pick one job, do it well, and make the model small enough that anyone can run it.
What to try this weekend
If you pay for transcription today, the honest test is a bake-off. Download Reson8/Rumble-2-Flash from Hugging Face, run an hour of your own audio through it, and compare the word error rate and the cost against whatever you use now. Pay attention to the languages and accents that actually show up in your data, not the ones in the press release. Multilingual benchmarks average over languages, and your worst language is the one that decides whether the model works for you.
There are limits worth knowing before you commit. The model is built for prerecorded audio, not live streaming, so real-time captioning is a different problem. Reson8 has not published memory or VRAM requirements, so claims about running it on a phone rest on third-party posts rather than the company's own documentation. And a 0.5B model will not match a 30B model on every accent or every noisy recording. The question is whether it is close enough for your use case, and at this size, for a lot of use cases, it will be.
Reson8 raised a 5 million euro pre-seed round in March, led by Balderton Capital, and the open-weights release is a bet that developers and community matter more than lock-in. It is a good bet. The speech market has spent years waiting for a small, open, multilingual model that is actually good. On November 18, it got one.


