model release

Reson8 releases 500M-parameter Rumble-2-Flash speech model as free Apache 2.0 open weights

TL;DR

Amsterdam-based Reson8 (Resonate Labs B.V.) has released Rumble-2-Flash, a 500 million-parameter transcription model covering more than 40 languages, as a free Apache 2.0 open-weights download. The company claims a 7.09% average word error rate across seven languages on Hugging Face's Open ASR Leaderboard multilingual tests.

3 min read
0

Amsterdam-based speech recognition startup Resonate Labs B.V., which operates as Reson8, today released Rumble-2-Flash as a free open-weights download under an Apache 2.0 license. The 500 million-parameter transcription model supports more than 40 languages and is designed for fast transcription of prerecorded audio.

Key specifications

  • Parameters: 500 million
  • Languages: more than 40
  • License: Apache 2.0
  • Use case: fast transcription of prerecorded audio
  • API pricing: from €0.01 per hour of audio

Rumble-2-Flash first appeared in Reson8's API on Nov. 3, a week before the company launched its flagship Rumble-2 model. Today's release adds the open-weights download.

Benchmark claims

Reson8 describes Rumble-2-Flash as one of the smallest models on Hugging Face's Open ASR Leaderboard. All figures below come from the company's own runs of the leaderboard's tests:

  • Multilingual average: a 7.09% word error rate across seven languages, including German, French and Dutch.
  • English: an average word error rate of about 5%, according to the company.
  • Lower-resource languages: a mean word error rate of 17.6% on a separate test covering 10 languages such as Maltese and Basque. Reson8 says OpenAI's Whisper Large V3 scored 39.82% on the same test.

Reson8 also says its models are trained for noisy conditions, including phone-line audio. "The real world doesn't happen in a recording studio," said Thomas Kluiters, the company's head of AI.

Positioning

Reson8 is targeting developers and researchers who need speech recognition for languages other than English. The company claims leading speech models are trained primarily on English, even though only about one in 20 people speak it as a first language, and says accuracy falls furthest on specialist terms such as medication names outside English. "Generic speech models are built for English and make noticeably more mistakes in other languages," said co-founder and CEO Raoul Ritter. "Languages like Dutch, Swedish and German are spoken by millions, yet they're an afterthought."

The company says its models are in production use for meeting notetakers, AI receptionists, customer support voice agents, medical transcription and captioning.

Rumble-2-Flash versus Rumble-2

Custom vocabulary support remains exclusive to Rumble-2. It lets customers add up to 50,000 terms as text without fine-tuning.

Hosting and compliance

Reson8's hosted service runs on the company's own GPUs in the Netherlands. The company says it retains no audio or transcripts. It also claims ISO 27001 certification and compliance with the EU's General Data Protection Regulation and AI Act.

Funding

Reson8 raised €5 million ($5.5 million) in pre-seed funding in March. Balderton Capital led the round, with NP-Hard Ventures participating. Balderton partner James Wise said speech is "quickly becoming a default interface across software."

What this means

Developers who need non-English transcription can now run a 500 million-parameter model on their own infrastructure under Apache 2.0, with no licensing fee. The hosted API is the alternative route, starting at €0.01 per hour of audio. The benchmark numbers are Reson8's own runs and have not been independently verified, including the Whisper Large V3 comparison on the 10 lower-resource languages. Teams with specialist vocabularies should note that custom vocabulary is available only in Rumble-2, not in the open-weights Rumble-2-Flash.

Related Articles

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use

Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

model release

Tencent releases Hy4 Flash Base: 220B-parameter MoE, 12B active, 1M-token context

Tencent has published Hy4 Flash Base on Hugging Face: a 220B-parameter mixture-of-experts model with 12B active parameters per token, text and image input, and a 1M-token context. Tencent calls it its most cost-effective Hy model to date.

Comments

Loading...