SigLino MoE 0.15B active
SigLino MoE 0.15B active routes each input through two of 28 experts. The card reports 0.15B active parameters and 0.6B total parameters. It is a sparse image-feature model.
How this encoder is configured
The encoder has 18 layers, a hidden width of 768 and 16-by-16 patches. It is distilled from DINOv3 and SigLIP2 and released under Apache 2.0. The active count describes routed computation, while the total describes the stored model.SigLino MoE 0.15B active model card
Evaluate the downstream feature task
Practical guidance. Compare runtime memory and throughput with the dense 0.6B model. Do not size the weight download using the active count. A sparse design can reduce computation without reducing the number of stored expert weights to the same degree.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/siglino-moe-0.15-0.6B |
| License | Apache 2.0SigLino MoE 0.15B active model card |
| Inputs and outputs | Images in; visual feature vectors outSigLino MoE 0.15B active model card |
Related models
A 0.6B dense vision encoder that learns image representations from DINOv3 and SigLIP2.
Launch date not confirmedDownloadableA sparse vision encoder with 0.6B total and about 0.3B active parameters that selects six of 28 experts.
Launch date not confirmedDownloadable