A Dutch startup gives away its speech-to-text model, and sells it as a service from one euro cent an hour
Reson8, a speech recognition startup in Amsterdam, has made its smallest transcription model free to download. The model, Rumble-2-Flash, converts recorded audio into text in more than 40 languages.
It is released under Apache 2.0, a permissive license. That means companies can use it in commercial products, modify it and run it on their own servers without paying Reson8. The model's public download page lists the same license.
For buyers of transcription, the move adds one more free option to a market where prices are already falling fast. For investors, it is a test of a familiar bet: give the basic product away and charge for the features businesses actually need.
What is free and what is not
Rumble-2-Flash has about 500 million parameters, the internal settings a model learns during training. That is small by current standards. Reson8 says it is roughly a third the size of the largest version of OpenAI's Whisper, a widely used free speech model.
Small matters for cost. A model this size can run on modest hardware, which cuts the bill for companies that transcribe thousands of hours of calls or meetings. The downloadable files are about 1 gigabyte, according to TokenPost's review of the listing.
Reson8 also sells the same model through its own hosted service. Prices start at €0.01 per hour of audio, so 1,000 hours costs about €10, the company says. That service went live on November 3; the free download is the new step.
The free model is deliberately limited. It only handles recordings, not live conversations. Reson8's own launch post says features such as telling speakers apart, live transcription, spotting when someone has finished talking, automatic language detection and custom models belong to its larger, paid Rumble-2 model.
One feature businesses often ask for, teaching the system up to 50,000 industry terms or product names, is also reserved for Rumble-2, SiliconANGLE reported. That split is the business model in plain view: the free model draws developers in, and the paid one serves the customers with specific needs.
There is also a small catch on access. The download page is gated, meaning users must log in and accept conditions before they can reach the files. Both the Hugging Face page and TokenPost note this. The page asks users to agree to share their Hugging Face account information.
How good is it? Mostly the company's word
Reson8's headline claims come from tests it ran itself. On multilingual tests drawn from Hugging Face's public speech-recognition leaderboard, it reports an average error rate of 7.09% across seven languages, including German, French and Dutch.
The bigger claim concerns less widely spoken languages. Across ten of them, such as Maltese and Basque, Reson8 says its model gets 17.6% of words wrong, against 39.8% for Whisper's largest version. Those are the company's own results.
The company says its own bars in comparison charts are self-tested, while some rival numbers come from others under different setups. That makes direct comparisons less reliable than the charts suggest.
One outside check exists. TypeWhisper, a free, open-source dictation app, published its own test of 980 audio clips in English, German and Russian on November 3. In that test, Rumble-2-Flash had the lowest English error rate of the cloud services measured, at 4.8%, and was the fastest.
But it also had the highest error rates in German and Russian among the services compared, at 7.6% and 10.8%. Soniox, a paid rival, scored 5.0% and 5.4% on the same languages with its latest real-time model. The test ran each clip once, so it is a snapshot, not a verdict.
That result cuts against Reson8's main pitch. The company argues generic models are built for English and stumble elsewhere. In at least one outside test, its own small model followed the same pattern.
Why a €5 million startup is giving its work away
Reson8 raised €5 million in pre-seed funding in March, led by London's Balderton Capital with NP-Hard Ventures. Its founders are chief executive Raoul Ritter, head of AI Thomas Kluiters and Jarno Verhagen.
That is very little money against the companies it compares itself with in its charts, including Deepgram, AssemblyAI, ElevenLabs, Nvidia and OpenAI. A free download is a cheap way to get noticed by developers who would never sit through a sales call.
Early uptake is modest. The model's page showed 444 downloads in the past month and 12 likes. Those numbers may reflect the gate and the short time since release, but they are far from the scale of established free models.
Reson8's second pitch is location. It runs its hosted service on its own hardware in the Netherlands and says it keeps no audio or transcripts. It also claims compliance with Europe's data protection rules and the EU AI Act. For hospitals, law firms and call centers in Europe, where the data is stored can decide a purchase.
Those compliance claims are the company's own, and buyers should ask for documents. Still, the combination is clear: a free model that customers can run inside their own walls, plus a cheap European-hosted service for those who would rather not.
The risk for Reson8 is that free speech models are multiplying, many from companies with far deeper pockets. Its free model has to be good enough to win attention, but not so complete that customers never pay for the bigger one. For buyers, the practical step is simple: test it on your own recordings, in your own languages, before believing anyone's chart.