MERaLiON 2 10B
MERaLiON 2 10B combines a localised Whisper speech encoder with Gemma 2 text generation. It handles several audio tasks through instructions. This is the general audio-understanding checkpoint, distinct from the separately tuned 10B ASR release.
A general audio instruction model
The card describes fine-tuning with 120,000 hours across six tasks. These include spoken questions, dialogue summaries and audio descriptions. The longer audio guidance applies to understanding tasks; it does not remove the shorter recommendation for transcription.MERaLiON 2 10B official model card
One model with alternate download formats
The official MLX and GGUF repositories provide alternate formats of the 10B release. They are grouped here because format conversion alone does not establish another trained model.MERaLiON official model catalog
The generation 2 license combines MIT, Whisper and Gemma terms with added conditions. It requires redistribution notices and acknowledgement for works developed using the model or outputs.MERaLiON Public Licence v3 for generation 2
Keep recognition and interpretation separate
Practical guidance. Use the general model when a question needs information from the audio beyond a transcript. For transcript-only work, compare the dedicated ASR checkpoint. Test short clips and longer recordings separately so duration does not hide a task-specific error.MERaLiON 2 10B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MERaLiON-2-10BMERaLiON-2-10B-MLXMERaLiON-2-10B-GGUF |
| Context window | 8,192 tokens Maximum text-decoder positions in the published configuration; audio consumes representation space and has separate duration guidance. MERaLiON 2 10B published configuration |
| Upstream components | OpenAI Whisper large-v3 and Google Gemma 2 9B IT.MERaLiON 2 10B published configuration |
| Text decoder | 42 layers with 3,584 hidden dimensions.MERaLiON 2 10B published configuration |
| Audio preparation | 16 kHz mono; 30 seconds recommended for ASR and speech translation.MERaLiON 2 10B official model card |
| Other audio tasks | Audio understanding is tested up to 300 seconds.MERaLiON 2 10B official model card |
Related models
- MERaLiON 2 10B ASR
MERaLiON / A*STAR IAIC · MERaLiON 2 ASR
A generation 2 speech recognizer tuned for English and Southeast Asian transcription.
Launch date not confirmedDownloadable - MERaLiON 2 3B
MERaLiON / A*STAR IAIC · MERaLiON 2
A compact audio-understanding model that combines Whisper large-v3 with a Gemma 2 2B IT text decoder.
Launch date not confirmedDownloadable - MERaLiON 3 10B
MERaLiON / A*STAR IAIC · MERaLiON 3
A third-generation speech-text model with regional fine-tuning for audio questions, dialogue summaries and contextual analysis.
Launch date not confirmedDownloadable