google/siglip-so400m-patch14-384: reference and serving guide
SigLIP so400m is the image-text embedding model we recommend by default for new multimodal search builds. It beats every CLIP checkpoint we serve on retrieval quality and costs less to run than the big one. The reasons are in the training, and there are two gotchas worth knowing before you deploy it.
Spec sheet
| Spec | Value |
|---|---|
| Embedding dimensions | 1,152 |
| Parameters | 878M |
| Inputs | text, image |
| Image resolution | 384 x 384 |
| Max text length | 64 tokens |
| Training data | WebLI |
| License | Apache 2.0 |
| Released | January 2024 (paper: arXiv 2303.15343) |
| Catalog page | superlinked.com/models/google-siglip-so400m-patch14-384 |
The odd-looking name is two design decisions. “so400m” is a shape-optimized Vision Transformer: instead of scaling a ViT the usual way, the authors used scaling-law work (arXiv 2305.13035) to pick compute-optimal width and depth at around 400M vision parameters. “patch14-384” is 14-pixel patches at 384-resolution input.
What sigmoid loss changes
CLIP trains with a softmax contrastive loss; every image-text pair is scored against every other pair in the batch, which requires a global view of the whole batch’s similarity matrix. SigLIP replaces that with a sigmoid loss over individual pairs. No global normalization. The practical consequences: training scales to larger batches, quality holds up better at small batch sizes, and the model’s raw outputs behave like independent probabilities per pair rather than scores that only mean something relative to a batch.
For you as a user, the loss function is history; the benchmark table is what remains. On Flickr30k image-to-text retrieval (NDCG@10, measured on an L4, same suite for every model in our catalog):
| Model | NDCG@10 | Cost $/1M tokens |
|---|---|---|
| SigLIP so400m | 0.900 | $0.493 |
| CLIP ViT-H/14 laion2B (986M) | 0.862 | $0.508 |
| CLIP ViT-L/14 openai (428M) | 0.782 | $0.227 |
| CLIP ViT-B/32 laion2B (151M) | 0.774 | $0.218 |
Higher quality than CLIP ViT-H with 108M fewer parameters and a lower cost per token. The small CLIPs remain the pick when serving cost dominates and quality tolerance is loose; nothing else about CLIP survives this table.
Two deployment gotchas
Tokenization is strict. SigLIP was trained with text padded to exactly 64 tokens, so you must tokenize with padding="max_length". Frameworks that pad dynamically to the longest sequence in a batch produce degraded similarities with no error message. If your text pipeline works with CLIP and quietly misbehaves with SigLIP, this is why. (SIE’s /encode endpoint applies the correct padding; the gotcha applies when you run the model yourself.)
Scoring is sigmoid, not softmax. If your downstream code softmaxes similarity scores across candidates, you are re-imposing the assumption SigLIP removed. Rank by raw score.
The 64-token text limit is tighter than CLIP’s 77 and cuts the same way: this is a model for captions and short queries, not paragraphs. Pair it with a text embedding model if you need both.
Serving numbers
From the catalog page, measured on an L4:
| Metric | Value |
|---|---|
| Latency (p50, headline) | 347 ms |
| Throughput | 451 tok/s headline; 700 tok/s corpus text, 7.3 megapixels/s images (b1 c16) |
| Cost | $0.493 per 1M tokens |
Image encoding is the expensive half; throughput for images is measured in megapixels per second, and a corpus of product photos costs meaningfully more to embed than its captions. Budget from the catalog numbers, not from text-model intuitions.
In SIE, SigLIP runs behind the same /encode endpoint as text models; send text or images, get 1,152-dim vectors in one shared space. Our wine recommender example uses exactly this pattern (photo in, OCR, match against inventory) if you want a working reference.
Where it sits in 2026
SigLIP 2 exists (google/siglip2-so400m-patch14-384, February 2025, also Apache 2.0), adding multilingual training and variants that handle native aspect ratios. We benchmark models before recommending them; check the catalog for current SigLIP 2 numbers before choosing between the generations. For English caption-style retrieval, the original so400m checkpoint’s numbers above are the ones to beat.
If your workload is document screenshots or UI images rather than photos, a different architecture wins: see Qwen3-VL-Embedding and ColQwen in the Qwen3 family guide.