Reson8 Open-Sources a Speech Model 60 Times Smaller Than NVIDIA's: 7.86% Word Error Rate vs. 7.78%
Transcribing calls usually means either a paid cloud API or a heavy model on an expensive GPU. Amsterdam-based startup Reson8 has released Rumble-2-Flash, a 0.5B-parameter model covering 40+ languages, under Apache 2.0. On AppTek's 302-hour test of call-center recordings, it gets 7.86% of words wrong, while NVIDIA's 30B-parameter Nemotron 3.5 Transcribe gets 7.78% wrong. The developer itself reported these figures, so run the model on 5–10 hours of your own recordings.
AI-processed from X: @curratedbrief; edited by Hamidun News
Imagine you need to turn thousands of hours of customer-support calls into text. There are usually two options: pay a cloud service for every minute of audio, or run a heavy model on an expensive GPU. Amsterdam-based startup Reson8 offers a third: Rumble-2-Flash, an open, compact speech recognition model. It has 0.5B parameters, and on a test of call-center recordings it gets 7.86% of words wrong. For comparison, NVIDIA's Nemotron 3.5 Transcribe, which has 30B parameters, gets 7.78% of words wrong. The difference is 0.08 percentage points, while the models differ in size by a factor of 60.
What exactly was released
Rumble-2-Flash is a speech-to-text model, meaning it converts speech into text. It supports more than 40 languages and is distributed under the Apache 2.0 license. For practitioners this detail matters: the license allows you to use the model in commercial products, fine-tune it, and embed it in closed-source services. You do not have to open-source your own code, and you need no separate agreement with the author.
Reson8 measured quality on the AppTek benchmark. It consists of 302 hours of call-center conversations, with noise, interruptions, and natural spoken language rather than clean studio read-aloud speech. The metric is WER (word error rate), the share of incorrectly recognized words. A WER of 7.86% means that out of every 100 spoken words, the model gets about eight wrong on average.
Why the comparison with NVIDIA stands out
NVIDIA's Nemotron 3.5 Transcribe scores a 7.78% WER on the same test. In practice the difference is barely noticeable: over 10,000 words, it amounts to about 8 words. Meanwhile, Rumble-2-Flash has 60 times fewer parameters.
Here is an everyday analogy: a small city car reaches the same destination in almost the same time as a truck with a huge engine, while burning noticeably less fuel. A smaller model usually needs less memory, runs faster, and fits on cheaper hardware. When you process thousands of hours of calls a day, this can noticeably change your infrastructure bill. That said, the original announcement contains no data on speed, memory use, or cost, so you will have to work out the savings yourself.
What to check before deployment
The result was published by the developer itself, and no independent measurements are visible yet. It is also a single benchmark covering a single type of speech. If your product involves podcasts, lectures, video calls with a poor microphone, or specialized vocabulary such as medicine or law, the numbers may turn out differently.
Another caveat: "40+ languages" does not mean equal quality in all of them. In models like this, languages with little training data usually suffer the most, and a good result in English guarantees nothing for a rare language or dialect. Finally, WER does not account for punctuation, speaker labels, or timestamps, and for meeting transcription these often matter more than a single wrong word.
What to do in practice
If you are building a voice interface, call analytics, or a transcription pipeline for RAG, Rumble-2-Flash deserves a spot on your list of quick tests. The plan is simple: take 5–10 hours of your own real recordings with reference transcripts, run them through this model and through the one you use now, and calculate WER on your own data. While you are at it, measure latency and memory use on your target hardware.
If the quality turns out to be comparable, the gain is control. You can run the model yourself, your customers' data does not leave for someone else's cloud, and the license creates no legal complications. For startups and teams with a limited GPU budget, this is a serious argument. If your recordings differ greatly from call-center calls, the claimed 7.86% remains a reference point, not a guarantee.
- Nov 19, 2026Kling Image 4.0 Renders in Native 8K and Generates 10-Frame Series from 2.33 ₽ per Image
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
- Nov 19, 2026Smysl.AI: An Agentic RAG Over Company Documents, Built in Kotlin Without Python
- Nov 19, 2026A chain of agents condensed 32 GB of company data into a single report, opening 30,000 tickets along the way
Key Questions Answered
What exactly was released?
Rumble-2-Flash is a speech-to-text model, meaning it converts speech into text. It supports more than 40 languages and is distributed under the Apache 2.0 license. For practitioners this detail matters: the license allows you to use the model in commercial products, fine-tune it, and embed it in closed-source services. You do not have to open-source your own code, and you need no separate agreement with the author. Reson8 measured quality on the AppTek benchmark.
Why the comparison with NVIDIA stands out?
NVIDIA's Nemotron 3.5 Transcribe scores a 7.78% WER on the same test. In practice the difference is barely noticeable: over 10,000 words, it amounts to about 8 words. Meanwhile, Rumble-2-Flash has 60 times fewer parameters. Here is an everyday analogy: a small city car reaches the same destination in almost the same time as a truck with a huge engine, while burning noticeably less fuel. A smaller model usually needs less memory, runs faster, and fits on cheaper hardware.
What to check before deployment?
The result was published by the developer itself, and no independent measurements are visible yet. It is also a single benchmark covering a single type of speech. If your product involves podcasts, lectures, video calls with a poor microphone, or specialized vocabulary such as medicine or law, the numbers may turn out differently. Another caveat: "40+ languages" does not mean equal quality in all of them.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.