MERaLiON 2 3B
MERaLiON 2 3B is the smaller general audio model in generation 2. It retains a Whisper large-v3 speech component while using Gemma 2 2B IT for text generation. It accepts audio and instructions and returns text.
A smaller language component
The 3B release belongs to the same audio instruction family as the 10B model. Its card describes the family training and tasks, but some training prose explicitly names 10B. This article does not turn that shared wording into an independently verified 3B training-data count.MERaLiON 2 3B official model card
The configuration preserves a 32-layer Whisper encoder. Most of the reduction is in the text decoder. The total system is therefore not simply a text-only 2B language model.MERaLiON 2 3B published configuration
Weights and format conversions
MLX and GGUF downloads are grouped with the parent 3B release. They do not have separate model articles.MERaLiON official model catalog
The published generation 2 license includes upstream Whisper and Gemma terms, notice duties and acknowledgement requirements. It is broader than a standalone MIT declaration.MERaLiON Public Licence v3 for generation 2
Evaluate the smaller decoder on the full task
Practical guidance. Compare this model with 10B on the same recordings and instructions. A shorter or cheaper generation is useful only if it preserves the required details. Check names, mixed-language speech and summary omissions separately from general answer fluency.MERaLiON 2 3B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MERaLiON-2-3BMERaLiON-2-3B-MLXMERaLiON-2-3B-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 3B published configuration |
| Text decoder | 26 layers with 2,304 hidden dimensions.MERaLiON 2 3B published configuration |
| Input | 16 kHz mono audio.MERaLiON 2 3B official model card |
| Duration guidance | 30 seconds for ASR and speech translation; up to 300 seconds tested for audio understanding.MERaLiON 2 3B 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