SigLino MoE 0.3B active
SigLino MoE 0.3B active selects six of 28 experts. Its published total is 0.6B parameters, with about 0.3B active. This is the less sparse of the two SigLino MoE releases.
How this encoder is configured
The model uses 18 layers, a hidden width of 768 and 16-by-16 patches. Its teachers are DINOv3 and SigLIP2. Both MoE variants share the same advertised total scale but use different numbers of active experts.SigLino MoE 0.3B active model card
Evaluate the downstream feature task
Practical guidance. Use the same feature head and retrieval queries to compare the two MoE variants. The additional active experts are useful only if their quality gain offsets the compute cost in the intended workload. Check performance with the actual image resolution.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/siglino-moe-0.3-0.6B |
| License | Apache 2.0SigLino MoE 0.3B active model card |
| Inputs and outputs | Images in; visual feature vectors outSigLino MoE 0.3B 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 0.15B active parameters that selects two of 28 experts.
Launch date not confirmedDownloadable