Runs as a custom pluginEmbeddingVerified 2026-11-12
DFN2B CLIP ViT-L/14
apple/DFN2B-CLIP-ViT-L-14
DFN2B CLIP ViT-L/14 is a image and text embedding model (apple-amlr, non-commercial). Can run on Mixpeek as an imported Hugging Face model.
Mixpeek does not run it today. It can run as a custom plugin: the weights load without remote code.
No demo for DFN2B CLIP ViT-L/14 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
- Published COCO 5k recall@1 59.6 and Flickr30k 84.7 at ViT-L size.
- Worst at
- Non-commercial license; ships no preprocessor_config.json (Mixpeek builds one from open_clip_config.json); 77-token text.
- Task
- Embedding
- Input
- Image, Text
- Output
- Vector
- Use cases
- Visual search
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
Image
- Resolution
- 224x224
- Color
- RGB
- Normalize
- CLIP mean (0.4815, 0.4578, 0.4082), std (0.2686, 0.2613, 0.2758)
Text
- Max tokens
- 77
Prompts
captions as written; CLIP BPE tokenizer (lowercases), 77 tokens
Output
Vector
- Dims
- 768
- 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 Imported from the Hugging Face Hub as an hf:// extractor and served by the transformers image-text loader, safetensors only, trust_remote_code false. GPU recommended at this size.
- Serve apps
- None
- Runtime
- transformers CLIPModel + CLIPProcessor
- Weights
- safetensors, no remote code
- Parameters
- 428 M
- Size on disk
- 1,710 MB
- Preprocessing
- Resize to 224x224 without a crop (open_clip resize_mode squash), CLIP mean/std normalize; tokenize to 77.
- Postprocessing
- L2-normalize; cosine similarity.
- 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 on Mixpeek yet; the SOTA bench measures it.
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is DFN2B CLIP ViT-L/14 best at?
- Published COCO 5k recall@1 59.6 and Flickr30k 84.7 at ViT-L size.
- What are the limits of DFN2B CLIP ViT-L/14?
- Non-commercial license; ships no preprocessor_config.json (Mixpeek builds one from open_clip_config.json); 77-token text.
- Can DFN2B CLIP ViT-L/14 be used commercially?
- No. Apple AMLR (sample code) license: research use; the weights are not licensed for commercial products.
- How does DFN2B CLIP ViT-L/14 run on Mixpeek?
- Imported from the Hugging Face Hub as an hf:// extractor and served by the transformers image-text loader, safetensors only, trust_remote_code false. GPU recommended at this size.
- What does DFN2B CLIP ViT-L/14 output?
- A 768-dimensional dense vector, compared by cosine similarity.