Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages

Amsterdam startup Reson8 released Rumble-2-Flash as open weights on Hugging Face: a 0.5-billion-parameter speech recognition model for 40+ languages under Apache 2.0 that the company says matches models up to 60 times larger on phone calls.

Christopher Ort
By Christopher Ort · 2 min read
Share
A microphone and headphones on a wooden desk beside a cassette recorder.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. How it works
  5. Why it matters
  6. What's still open
  7. Sources
  8. i10x Perspective

A ten-person startup from Amsterdam has released one of the smallest competitive speech recognition models as open weights. On 18 November 2026 Reson8 published Rumble-2-Flash on Hugging Face under the Apache 2.0 licence: a 0.5-billion-parameter speech-to-text model for more than 40 languages that the company says matches models many times its size.

i10X Quick Take

Quick Take

  1. The model: Rumble-2-Flash, 0.5B parameters, automatic speech recognition for 40+ languages with automatic language detection.
  2. Licence: Apache 2.0, with native support in Hugging Face Transformers and vLLM, no custom code needed.
  3. Speed: over 7,000 RTFx on full batches, about half a second for an hour of audio, according to Reson8.
  4. Phone calls: 7.86% word error rate on AppTek’s 302-hour call-centre test, versus 7.78% for NVIDIA’s 30B Nemotron 3.5 Transcribe, per Reson8’s own runs.
  5. API price: from €0.01 per hour of audio through Reson8’s hosted API.

In plain English

Speech-to-text models turn recordings into written words. Most accurate ones are big and expensive to run. Reson8 shrank its larger model into a small one that still gets about nine in ten words right in noisy meetings and phone calls, and now lets anyone download it for free and run it on their own servers.

What happened

Reson8 announced the open release in a Hugging Face blog post: “Today we’re open sourcing Rumble-2-Flash under the Apache 2.0 license. It’s a 0.5B-parameter speech recognition model for 40+ languages that competes with models 3-60x its size.” The model debuted in Reson8’s API on 3 November, a week before its larger flagship Rumble-2, according to SiliconANGLE.

How it works

  • Architecture: it uses the Qwen3-ASR architecture, which is why it runs in standard Transformers and vLLM.
  • Training: on-policy distillation from Rumble-2. The small model transcribes audio itself and the larger teacher grades its output token by token.
  • Output control: numbers as digits or words, optional filler words, verbatim repeats and forced language.
  • Benchmarks: Reson8 reports a 6.6% average English word error rate on the Open ASR Leaderboard and says it beats Meta’s omniASR-LLM 7B and Whisper large-v3 on the multilingual track. On ten lower-resource languages it reports 17.6% average WER against 39.8% for Whisper Large V3.
  • Missing by design: speaker diarization and live streaming stay in the larger Rumble-2.

Why it matters

For developers, a permissively licensed 0.5B model that runs in vLLM makes self-hosted transcription cheap enough for full call archives or on-device pipelines.

For companies in Europe, Reson8 says its hosted service runs on its own GPUs in the Netherlands and keeps no audio or transcripts, which speaks to GDPR concerns.

For the open model scene, small specialised models keep closing the gap with much larger systems such as NVIDIA’s Nemotron 3.5 Transcribe.

What’s still open

  • All benchmark numbers are Reson8’s own; independent results on the Open ASR Leaderboard and on real customer audio will matter more.
  • How it performs on overlapping speakers and very noisy audio compared with the larger Rumble-2.

Sources

  • Hugging Face blog, “Rumble-2-Flash: an open 0.5B speech model for 40+ languages,” 18 November 2026
  • Hugging Face model card, Reson8/Rumble-2-Flash
  • Reson8 blog, “Meet Rumble-2-Flash,” 3 November 2026
  • SiliconANGLE, report on the open-weights release, 18 November 2026

i10x Perspective

Speech recognition is becoming a commodity layer, and Reson8 is pushing it there faster with a free model and a one-cent-per-hour API. The real test is whether a team of ten can keep pace with NVIDIA, Alibaba and OpenAI on accuracy updates. More open releases in our Models section.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

ModelsNews

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests