- 0.5B parameters and about 1GB, enabling fully offline on-device transcription of calls and voice notes
- Covers 45 languages and is free under Apache 2.0, with open weights available on Hugging Face and elsewhere
- Optimized for real-world hard audio such as thick accents, overlapping speakers, bad phone lines and street noise
AI CURATED HOT STORY · 11/18/2026
Reson8 Open-Sources 1GB On-Device Speech Model Rumble-2-Flash
Amsterdam-based startup Reson8 has open-sourced Rumble-2-Flash, an on-device speech recognition model with only 0.5B parameters and roughly 1GB in size, under the Apache 2.0 license. Despite its compact footprint, it covers 45 languages with automatic language detection and is claimed to rival models 3 to 60 times larger. This release marks a significant step toward bringing high-quality multilingual speech recognition to offline mobile scenarios.
Technically, Rumble-2-Flash natively supports Hugging Face Transformers and vLLM, enabling deployment without custom code, and delivers impressive throughput—processing one hour of audio in about half a second. It offers output controls for number formatting, filler words, verbatim repetition, and forced language, and is optimized for challenging real-world audio such as heavy accents, overlapping speakers, poor phone lines, and street noise, directly addressing scenarios where traditional transcription tools fail.
Industry impact is substantial: by achieving multilingual, robust speech recognition with minimal resource requirements, the model lowers the barrier for developers to deploy speech capabilities on edge devices, accelerating adoption in offline voice assistants, real-time meeting transcription, and accessibility tools. The permissive Apache 2.0 license and open weights on Hugging Face are poised to spur innovation and competition in edge-based speech AI.
Hourly heat curve
43 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 2 reports · Latest first- Rohan Paul (@rohanpaul_ai)T2·72 pts
- Hugging Face BlogT1·72 pts
- 0.5B parameters cover 40+ languages with automatic language detection, claimed to rival models 3-60x larger
- Native support in Hugging Face Transformers and vLLM enables deployment without custom code
- High throughput processes an hour of audio in about half a second, with output controls for digit formatting and verbatim transcription