NEWVectors or files. Pick a path.Start →
    Served on MixpeekEmbeddingRecommendedVerified 2026-11-05

    SigLIP Base patch16 224

    google/siglip-base-patch16-224

    SigLIP Base patch16 224 is a image and text embedding model (Apache-2.0). Served on Mixpeek, scaling to zero when idle.

    A Ray Serve app in Mixpeek's inference registry hosts it. It starts at zero replicas and wakes on first use.

    Use in a new namespaceAdd to an existing namespace Record as JSON

    New opens the Designer in Studio. For an existing namespace, open it in Studio and use its Designer tab.

    No demo for SigLIP Base patch16 224 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
    Text-to-image and image-to-image search over photos, products and video frames in one shared space.
    Worst at
    Reading text in images, counting, small objects, and catalog facts that are not visible (NGA eval, self-portrait and etching queries).
    Task
    Embedding
    Input
    Image, Text
    Output
    Vector
    Use cases
    Visual search, Product search
    Why we recommend it
    The served image embedding behind image_extractor; measured on the NGA art search eval.

    Templates that use it

    • Video Moderation Extraction Foundation

      Segment a video corpus, extract a keyframe record for every segment, and match brand marks against a reference set you upload. Face and public-figure matching run on a dedicated or customer-hosted deployment. This deploy is the extraction layer a moderation policy sits on, and the page carries a state on every part so you can see which policy pieces are yours to add.

      Deploy this template

    Input and output

    Input

    image
    resolution
    224x224
    color
    RGB
    normalize
    mean 0.5, std 0.5
    text
    max_tokens
    64
    padding
    max_length
    language
    English
    prompts
    plain captions; no instruction prefix

    Output

    vector
    dims
    768
    distance
    cosine
    type
    dense

    How Mixpeek wires it

    Extractors
    • image_extractor@v1 fixed
    • web_scraper@v1 generate_image_embeddings=true (default true)
    Feature URIs
    • mixpeek://image_extractor@v1/google_siglip_base_v1
    • mixpeek://web_scraper@v1/google__siglip_base_patch16_224
    Vector indexes
    image_extractor_v1_embeddinggoogle__siglip_base_patch16_224
    Retriever stages
    feature_search, mmr
    Combines with
    • image_extractor: one SigLIP vector per image; feature_search embeds a text or image query with the same Serve app
    • web_scraper: SigLIP vectors for page images beside E5 text, Jina code and DINOv2 structure vectors
    • colqwen_extractor reuses the image_extractor batch step with its embed swapped for ColQwen2

    How it runs

    Status
    Served on Mixpeek Ray Serve app in the inference registry. image_extractor batch calls the Serve app (ImageEmbeddingBatchService, strict_serve: no local fallback unless SERVE_PROXY_ALLOW_LOCAL_FALLBACK=1); web_scraper loads SiglipVisionModel in-process.
    Serve apps
    google__siglip_base_patch16_224
    Runtime
    transformers AutoModel + AutoProcessor (SiglipModel) for realtime text and image; SiglipVisionModel + SiglipImageProcessor for batch (SiglipImageEmbeddingBatch, web_scraper DocTypeImageEmbedder)
    Weights
    safetensors, no remote code
    Parameters
    203 M
    Size on disk
    813 MB
    Preprocessing
    Resize to 224x224, rescale, normalize mean 0.5 std 0.5; text padded to 64 tokens.
    Postprocessing
    L2-normalize the pooled image or text embedding (pooler_output).
    Serving
    both, batch 64
    Hardware
    Runs on CPU (slower without a GPU)

    Scale to zero

    Minimum replicas
    0
    Maximum replicas
    4
    Idle timeout
    300 s
    Memory
    1 GB
    CPUs
    0.5
    GPUs
    Set at deploy time: hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0
    Cold start, node warm
    Not recorded
    Cold start, new node
    Not recorded

    Measured cost

    Unmeasured

    No committed throughput or cold-start measurement for this model alone.

    Our evaluations

    Dataset and metricScoreBaselineDate
    NGA art search, 25 text queries
    P@10 (regex relevance, lower bound) (k=10)
    canvas/apps/nga-art-search/eval/RESULTS.md
    		0.748None2026-11-02
    NGA art search, 25 text queries
    nDCG@10 (k=10)
    canvas/apps/nga-art-search/eval/RESULTS.md
    		0.738None2026-11-02
    ViDoRe v1 + v2 document pages, 1,500 English queries (93,992-item mixed corpus)
    nDCG@10, SigLIP alone top 10 (dense single vector) (k=10)
    research/goal-specs/mvs-scale-from-zero/RESULTS-MULTIMODAL-LI.md
    		0.135exact ColQwen2 MaxSim: 0.6962026-10-15
    MSR-VTT 1K-A captions over scene-cut keyframes (1,000 queries)
    nDCG@10, SigLIP alone top 10 (k=10)
    research/goal-specs/mvs-scale-from-zero/RESULTS-MULTIMODAL-LI.md
    		0.14exact ColQwen2 MaxSim: 0.1332026-10-15
    Marqo/KAGL product images, 1,000 product-title queries
    nDCG@10, SigLIP alone top 10 (k=10)
    research/goal-specs/mvs-scale-from-zero/RESULTS-MULTIMODAL-LI.md
    		0.293exact ColQwen2 MaxSim: 0.172026-10-15

    Questions

    What is SigLIP Base patch16 224 best at?
    Text-to-image and image-to-image search over photos, products and video frames in one shared space.
    What are the limits of SigLIP Base patch16 224?
    Reading text in images, counting, small objects, and catalog facts that are not visible (NGA eval, self-portrait and etching queries).
    Can SigLIP Base patch16 224 be used commercially?
    Yes. Apache-2.0 weights and code; no use restrictions.
    How does SigLIP Base patch16 224 run on Mixpeek?
    Ray Serve app in the inference registry. image_extractor batch calls the Serve app (ImageEmbeddingBatchService, strict_serve: no local fallback unless SERVE_PROXY_ALLOW_LOCAL_FALLBACK=1); web_scraper loads SiglipVisionModel in-process.