MERaLiON 3 10B
MERaLiON 3 10B is the general speech-text model in the third generation. It develops the Whisper and Gemma audio-language architecture through regional fine-tuning. It answers audio questions and generates text, while the separate ASR line focuses on transcription.
Speech, context and audio descriptions
The six training tasks cover recognition, spoken questions, dialogue summaries, captions, audio-scene questions and contextual paralinguistic questions. These outputs are generated interpretations of audio. A fluent answer does not verify that the event or attribute was correctly inferred.MERaLiON 3 10B official model card
Downloadable generation 3 weights
The public weights use the 2026 MERaLiON 3 license. Its component terms and acknowledgement requirements apply alongside the custom redistribution conditions.MERaLiON 3 Public Licence, 2026
The public audio-chat service does not establish an exact checkpoint mapping in its visible interface. This article therefore confirms weight access without assigning every hosted chat feature to this 10B release.MERaLiON developer API console
Check grounded answers against the recording
Practical guidance. Ask for observable details before asking for a broad interpretation. For example, evaluate whether the model reports the right words and sounds before scoring a dialogue summary. Inspect omissions and unsupported additions separately.MERaLiON 3 10B published configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MERaLiON-3-10B |
| Context window | 8,192 tokens Maximum text-decoder positions in the published configuration; audio consumes representation space and has separate duration guidance. MERaLiON 3 10B published configuration |
| Upstream models | Whisper large-v3 and Gemma 2 9B IT, with local adaptation.MERaLiON 3 10B official model card |
| Text decoder | 42 layers; 3,584 hidden dimensions.MERaLiON 3 10B published configuration |
| Fine-tuning scale | 150,000 hours across six tasks.MERaLiON 3 10B official model card |
| Audio guidance | 16 kHz mono; 30 seconds recommended for ASR and speech translation, with 300 seconds tested for general audio understanding.MERaLiON 3 10B official model card |
Related models
- MERaLiON 2 10B
MERaLiON / A*STAR IAIC · MERaLiON 2
A general audio-understanding model that combines a localized Whisper encoder with Gemma 2 text generation.
Launch date not confirmedDownloadable - MERaLiON 3 ASR
MERaLiON / A*STAR IAIC · MERaLiON 3 ASR
A regional speech recognition release with a public 3B-class checkpoint and a separately described hosted service.
Launch date not confirmedDownloadable · API · Provider product - MERaLiON SER v1
MERaLiON / A*STAR IAIC · MERaLiON Speech Emotion Recognition
A 309M speech classifier that uses Whisper-Medium features to estimate emotions and continuous affect values.
Launch date not confirmedDownloadable · API · Provider product