Runs as a custom pluginEmbeddingVerified 2026-11-05
SigLIP so400m patch14 (224/384)
google/siglip-so400m-patch14-384
SigLIP so400m patch14 (224/384) is a image and text embedding model (Apache-2.0). Can run on Mixpeek as a custom plugin.
Mixpeek does not run it today. It can run as a custom plugin: the weights load without remote code.
No demo for SigLIP so400m patch14 (224/384) yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Strongest SigLIP v1 image-text model; the vision tower behind many VLMs (PaliGemma).
- Worst at
- About 0.9B parameters: GPU in practice for video frames; 1152-d vectors.
- Task
- Embedding
- Input
- Image, Text
- Output
- Vector
- Use cases
- Visual search, Product search
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- image
- resolution
- 384x384 (224 sibling)
- color
- RGB
- normalize
- mean 0.5, std 0.5
- text
- max_tokens
- 64
- prompts
- short captions; the card uses 'This is a photo of {label}.'; text padded to max_length 64
Output
- vector
- dims
- 1152
- distance
- cosine
- type
- dense
- note
- image and text vectors share one space; L2-normalize before cosine
How Mixpeek wires it
- Extractors
- None today
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- feature_search, mmr
- Combines with
- video frames (VideoFrameExpander) -> image vectors -> feature_search with a text query
- detector boxes (RT-DETR, D-FINE) -> crops -> image vectors per object
- image vectors + E5 text vectors on captions -> feature_search fusion (dbsf)
How it runs
- Status
- Runs as a custom plugin Not served. Would run as a custom extractor plugin (engine/extractors/<name>/v1/ with realtime.py for query embedding), loading safetensors with trust_remote_code false. GPU recommended at this size.
- Serve apps
- None
- Runtime
- transformers AutoModel + AutoProcessor (SiglipModel)
- Weights
- safetensors, no remote code
- Parameters
- 878 M
- Size on disk
- 3,512 MB
- Preprocessing
- Resize to 384x384, rescale to 0..1, normalize mean 0.5 std 0.5; text padded to 64 tokens (max_length).
- Postprocessing
- L2-normalize pooled image and text embeddings; sigmoid of scaled logits for zero-shot scores.
- Serving
- both, batch 32
- Hardware
- Runs on CPU (slower without a GPU)
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 1
- Idle timeout
- 600 s
- Memory
- 6 GB
- CPUs
- 2
- GPUs
- 1
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed measurement; plan row 3 measures it in the standalone image.
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is SigLIP so400m patch14 (224/384) best at?
- Strongest SigLIP v1 image-text model; the vision tower behind many VLMs (PaliGemma).
- What are the limits of SigLIP so400m patch14 (224/384)?
- About 0.9B parameters: GPU in practice for video frames; 1152-d vectors.
- Can SigLIP so400m patch14 (224/384) be used commercially?
- Yes. Apache-2.0 weights and code; no use restrictions.
- How does SigLIP so400m patch14 (224/384) run on Mixpeek?
- Not served. Would run as a custom extractor plugin (engine/extractors/<name>/v1/ with realtime.py for query embedding), loading safetensors with trust_remote_code false. GPU recommended at this size.
- What does SigLIP so400m patch14 (224/384) output?
- A 1152-dimensional dense vector, compared by cosine similarity.