Skip to main content

Searching

Immich uses Postgres as its search database for both metadata and contextual CLIP search.

Contextual CLIP search is powered by the pgvecto.rs extension, utilizing machine learning models like CLIP to provide relevant search results. This allows for freeform searches without requiring specific keywords in the image or video metadata.

Advanced Search Filters

In addition, Immich offers advanced search functionality, allowing you to find specific content using customizable search filters. These filters include location, one or more faces, specific albums, and more. You can try out the search filters on the Demo site.

The filters smart search allows you to search by include:

  • People
  • Location
    • Country
    • State
    • City
  • Camera
    • Make
    • Model
  • Date range
  • File name or extension
  • Media type
    • Image (including live/motion photos)
    • Video
    • All
  • Condition
    • Not in any album
    • Archived
    • Favorited
    • Rating

Some search examples:

Configuration

Navigating to Administration > Settings > Machine Learning Settings > Smart Search will show the options available.

CLIP models

The default search model is fast, but there are many other options that can provide better search results. The tradeoff of using these models is that they're slower and/or use more memory (both when indexing images with background Smart Search jobs and when searching).

The first step of choosing the right model for you is to know which languages your users will search in.

If your users will only search in English, then the CLIP section is the first place to look. This is a curated list of the models that generally perform the best for their size class. The models here are ordered from higher to lower quality. This means that the top models will generally rank the most relevant results higher and have a higher capacity to understand descriptive, detailed, and/or niche queries. The models are also generally ordered from larger to smaller, so consider the impact on memory usage, job processing and search speed when deciding on one. The smaller models in this list are not too different in quality and many times faster.

Multilingual models are also available so users can search in their native language. Use these models if you expect non-English searches to be common. They can be separated into two search patterns:

  • nllb models expect the search query to be in the language specified in the user settings
  • xlm and siglip2 models understand search text regardless of the current language setting

nllb models tend to perform the best and are recommended when users primarily searches in their native, non-English language. xlm and siglip2 models are more flexible and are recommended for mixed language search, where the same user might search in different languages at different times.

For more details, check the tables below to see how they compare in memory usage, speed and quality by language.

Once you've chosen a model, follow these steps:

  1. Copy the name of the model (e.g. ViT-B-16-SigLIP__webli)
  2. Go to the Smart Search settings
  3. Paste the model name into the Model Name section
  4. Save the settings
  5. Go to the Job Status page
  6. Click "All" next to "Smart Search" to begin re-processing your assets with the new model
  7. (Optional) Confirm that the logs for the server and machine learning service don't have relevant errors

In rare instances, changing the model might leave bits of the old model's incompatible data in the database, causing errors when processing Smart Search jobs. If you notice errors like this in the logs, you can change the model back to the previous one and save, then repeat steps 3-7.

Please note that memory and execution time values are only estimates: actual usage will be different depending on many factors. As such, it's mainly intended as a way to compare the relative tradeoffs of each model.

Reference

Memory and execution time estimates were obtained without acceleration on a 7800x3D processor running bare metal Linux. All testing and evaluation was done at f32 precision (the default in Immich).

Execution Time (ms): After warming up the model with one pass, the mean execution time of 100 passes with the same input.

Memory (MiB): The peak RSS usage of the process afer performing the above timing benchmark. Does not include image decoding, concurrent processing, the web server, etc., which are relatively constant factors.

Recall (%): Evaluated on Crossmodal-3600, the average of the recall@1, recall@5 and recall@10 results for zeroshot image retrieval. Chinese (Simplified), English, French, German, Italian, Japanese, Korean, Polish, Russian, Spanish and Turkish are additionally tested on XTD-10. Chinese (Simplified) and English are additionally tested on Flickr30k. The recall metrics are the average across all tested datasets.

Pareto Optimal: Whether the model is not completely outclassed by another model. Try to use models that are optimal for the languages relevant to you. Specifically, for a given model and language, if there's another model that's better for that language in at least one respect (memory usage, execution time, recall) while being at least as good for that language in every other way, then the model is not optimal for that language.


English
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-384__webli385456.5785.99
ViT-SO400M-14-SigLIP2-378__webli394072.2585.96
ViT-gopt-16-SigLIP2-384__webli6585146.8485.96
ViT-SO400M-16-SigLIP2-512__webli4050107.6785.93
ViT-H-14-378-quickgelu__dfn5b5049108.485.78
ViT-L-16-SigLIP2-512__webli335892.5985.75
ViT-SO400M-16-SigLIP2-256__webli361127.8485.62
ViT-SO400M-14-SigLIP2__webli362227.6385.53
ViT-gopt-16-SigLIP2-256__webli647564.5185.48
ViT-L-16-SigLIP2-384__webli305751.785.47
ViT-H-14-quickgelu__dfn5b470138.7485.09
ViT-L-16-SigLIP2-256__webli283023.7785.03
ViT-B-16-SigLIP2__webli30385.8184.86
ViT-SO400M-14-SigLIP-384__webli441772.1984.61
ViT-L-16-SigLIP-384__webli339647.684.17
ViT-L-16-SigLIP-256__webli316023.8483.51
ViT-B-16-SigLIP-512__webli182826.1783.28
nllb-clip-large-siglip__v1422675.0583.24
nllb-clip-large-siglip__mrl424875.4483.23
ViT-B-16-SigLIP-384__webli112813.5383.19
ViT-L-14-quickgelu__dfn2b221220.4982.54
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1482.43
ViT-H-14__laion2b-s32b-b79k467639.0682.36
ViT-B-32-SigLIP2-256__webli30613.3182.28
ViT-B-16-SigLIP__webli10815.7781.9
ViT-B-16-SigLIP-256__webli11027.1181.9
ViT-L-14__laion2b-s32b-b82k223320.5680.82
nllb-clip-base-siglip__mrl469616.9580.65
nllb-clip-base-siglip__v1467515.1780.16
ViT-B-16-SigLIP-i18n-256__webli30296.8779.78
ViT-L-14__laion400m_e31218319.8778.64
ViT-L-14__laion400m_e32221819.7378.6
ViT-B-16-plus-240__laion400m_e3212466.9578.06
ViT-B-16-plus-240__laion400m_e3112636.9478.06
ViT-B-32__laion2b-s34b-b79k10012.2977.62
ViT-B-32__laion2b_e1610042.3877.47
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.276.91
ViT-B-16__laion400m_e329754.9876.43
ViT-B-16__laion400m_e319915.0476.35
ViT-B-32__laion400m_e319992.2873.83
ViT-B-32__laion400m_e3210032.3573.62
RN50x64__openai507948.7973.34
ViT-L-14__openai221219.9172.99
ViT-L-14-336__openai261643.4572.76
RN50x16__openai222115.8772.59
RN50x4__openai14165.8570.8
ViT-B-16__openai9855.0370.01
ViT-B-32__openai10042.2669.9
RN101__openai11113.2169.3
RN50__openai9132.3969.02
RN50__cc12m9142.3764.59
RN101__yfcc15m11113.2255.21
RN50__yfcc15m9082.3453.63
Arabic
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4477.3
nllb-clip-large-siglip__v1422675.0576.44
nllb-clip-base-siglip__mrl469616.9574.03
nllb-clip-base-siglip__v1467515.1773.19
ViT-SO400M-16-SigLIP2-384__webli385456.5769.31
ViT-SO400M-14-SigLIP2-378__webli394072.2569.29
ViT-SO400M-16-SigLIP2-512__webli4050107.6769.29
ViT-SO400M-16-SigLIP2-256__webli361127.8468.64
ViT-L-16-SigLIP2-512__webli335892.5968.35
ViT-L-16-SigLIP2-384__webli305751.768.25
ViT-SO400M-14-SigLIP2__webli362227.6368.23
ViT-gopt-16-SigLIP2-384__webli6585146.8467.56
ViT-gopt-16-SigLIP2-256__webli647564.5167.28
ViT-L-16-SigLIP2-256__webli283023.7766.89
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1466.52
ViT-B-16-SigLIP-i18n-256__webli30296.8764.1
ViT-B-16-SigLIP2__webli30385.8161.71
ViT-B-32-SigLIP2-256__webli30613.3160.7
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.259.66
Bengali
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0576.16
nllb-clip-large-siglip__mrl424875.4475.83
nllb-clip-base-siglip__mrl469616.9573.75
nllb-clip-base-siglip__v1467515.1773.34
ViT-B-16-SigLIP-i18n-256__webli30296.8736.43
ViT-SO400M-14-SigLIP2__webli362227.6326.56
ViT-SO400M-16-SigLIP2-256__webli361127.8426.54
ViT-SO400M-16-SigLIP2-384__webli385456.5726.19
ViT-SO400M-14-SigLIP2-378__webli394072.2526.19
ViT-SO400M-16-SigLIP2-512__webli4050107.6725.92
ViT-gopt-16-SigLIP2-384__webli6585146.8425.15
ViT-gopt-16-SigLIP2-256__webli647564.5124.18
ViT-L-16-SigLIP2-384__webli305751.721.44
ViT-L-16-SigLIP2-512__webli335892.5921.11
ViT-L-16-SigLIP2-256__webli283023.7720.94
Chinese (Simplified)
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0579.7
nllb-clip-large-siglip__mrl424875.4478.94
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1475.22
nllb-clip-base-siglip__v1467515.1774.8
nllb-clip-base-siglip__mrl469616.9573.91
ViT-gopt-16-SigLIP2-384__webli6585146.8472.8
ViT-SO400M-16-SigLIP2-512__webli4050107.6772.77
ViT-SO400M-14-SigLIP2-378__webli394072.2572.41
ViT-SO400M-16-SigLIP2-384__webli385456.5772.36
ViT-gopt-16-SigLIP2-256__webli647564.5171.59
ViT-L-16-SigLIP2-512__webli335892.5971.37
ViT-SO400M-16-SigLIP2-256__webli361127.8471.3
ViT-L-16-SigLIP2-384__webli305751.771.11
ViT-SO400M-14-SigLIP2__webli362227.6370.95
ViT-L-16-SigLIP2-256__webli283023.7770.51
ViT-B-16-SigLIP-i18n-256__webli30296.8767.48
ViT-B-16-SigLIP2__webli30385.8166.84
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.265.7
ViT-B-32-SigLIP2-256__webli30613.3163.38
Croatian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4487.46
nllb-clip-large-siglip__v1422675.0587.19
nllb-clip-base-siglip__mrl469616.9582.98
nllb-clip-base-siglip__v1467515.1782.92
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1481.93
ViT-SO400M-14-SigLIP2-378__webli394072.2573.77
ViT-SO400M-16-SigLIP2-512__webli4050107.6773.21
ViT-SO400M-16-SigLIP2-384__webli385456.5773.2
ViT-gopt-16-SigLIP2-256__webli647564.5172.95
ViT-SO400M-16-SigLIP2-256__webli361127.8472.89
ViT-gopt-16-SigLIP2-384__webli6585146.8472.88
ViT-SO400M-14-SigLIP2__webli362227.6372.85
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.272.69
ViT-L-16-SigLIP2-512__webli335892.5970.73
ViT-B-16-SigLIP-i18n-256__webli30296.8770.45
ViT-L-16-SigLIP2-384__webli305751.770.43
ViT-L-16-SigLIP2-256__webli283023.7769.97
ViT-B-16-SigLIP2__webli30385.8154.31
ViT-B-32-SigLIP2-256__webli30613.3153.3
ViT-H-14-378-quickgelu__dfn5b5049108.435.64
ViT-H-14-quickgelu__dfn5b470138.7435.17
ViT-L-16-SigLIP-256__webli316023.8433.65
ViT-L-16-SigLIP-384__webli339647.633.55
ViT-B-16-SigLIP-256__webli11027.1120.05
Cusco Quechua
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4438.08
nllb-clip-large-siglip__v1422675.0537.87
nllb-clip-base-siglip__mrl469616.9533.41
nllb-clip-base-siglip__v1467515.1733.06
Czech
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4473.76
nllb-clip-large-siglip__v1422675.0571.57
nllb-clip-base-siglip__mrl469616.9569.86
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1467.49
nllb-clip-base-siglip__v1467515.1767.15
ViT-gopt-16-SigLIP2-384__webli6585146.8463.62
ViT-SO400M-14-SigLIP2-378__webli394072.2563.35
ViT-gopt-16-SigLIP2-256__webli647564.5163.09
ViT-SO400M-16-SigLIP2-512__webli4050107.6763.07
ViT-SO400M-16-SigLIP2-384__webli385456.5762.98
ViT-SO400M-16-SigLIP2-256__webli361127.8462.82
ViT-SO400M-14-SigLIP2__webli362227.6362.73
ViT-L-16-SigLIP2-512__webli335892.5962.29
ViT-L-16-SigLIP2-384__webli305751.762.12
ViT-L-16-SigLIP2-256__webli283023.7761.74
ViT-B-16-SigLIP-i18n-256__webli30296.8761.52
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.261.01
ViT-B-16-SigLIP2__webli30385.8154.81
ViT-B-32-SigLIP2-256__webli30613.3154.31
ViT-L-16-SigLIP-256__webli316023.8433.58
ViT-L-16-SigLIP-384__webli339647.633.48
ViT-H-14-378-quickgelu__dfn5b5049108.432.38
ViT-H-14-quickgelu__dfn5b470138.7432.32
ViT-B-16-SigLIP__webli10815.7722.89
ViT-B-16-SigLIP-512__webli182826.1722.66
ViT-B-16-SigLIP-256__webli11027.1122.6
ViT-B-16-SigLIP-384__webli112813.5322.25
Danish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0587.16
nllb-clip-large-siglip__mrl424875.4486.88
nllb-clip-base-siglip__mrl469616.9584.18
nllb-clip-base-siglip__v1467515.1784.03
ViT-gopt-16-SigLIP2-384__webli6585146.8483.75
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1483.32
ViT-gopt-16-SigLIP2-256__webli647564.5183.25
ViT-SO400M-16-SigLIP2-384__webli385456.5782.3
ViT-SO400M-14-SigLIP2-378__webli394072.2582.19
ViT-SO400M-16-SigLIP2-512__webli4050107.6781.87
ViT-SO400M-14-SigLIP2__webli362227.6381.44
ViT-SO400M-16-SigLIP2-256__webli361127.8481.42
ViT-L-16-SigLIP2-512__webli335892.5980.0
ViT-L-16-SigLIP2-384__webli305751.779.82
ViT-L-16-SigLIP2-256__webli283023.7779.08
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.275.07
ViT-B-16-SigLIP-i18n-256__webli30296.8774.84
ViT-B-16-SigLIP2__webli30385.8167.68
ViT-B-32-SigLIP2-256__webli30613.3167.2
ViT-H-14-quickgelu__dfn5b470138.7465.59
ViT-H-14-378-quickgelu__dfn5b5049108.465.36
ViT-L-14-quickgelu__dfn2b221220.4942.31
ViT-L-16-SigLIP-256__webli316023.8441.46
ViT-L-16-SigLIP-384__webli339647.640.52
ViT-B-16-SigLIP-512__webli182826.1731.31
ViT-B-16-SigLIP-256__webli11027.1130.97
ViT-B-16-SigLIP__webli10815.7730.87
ViT-B-16-SigLIP-384__webli112813.5330.51
Dutch
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-512__webli4050107.6780.05
ViT-gopt-16-SigLIP2-384__webli6585146.8479.81
ViT-SO400M-16-SigLIP2-384__webli385456.5779.72
ViT-SO400M-14-SigLIP2-378__webli394072.2579.72
ViT-L-16-SigLIP2-512__webli335892.5979.64
ViT-L-16-SigLIP2-384__webli305751.779.49
nllb-clip-large-siglip__mrl424875.4479.41
nllb-clip-large-siglip__v1422675.0579.31
ViT-SO400M-16-SigLIP2-256__webli361127.8478.92
ViT-SO400M-14-SigLIP2__webli362227.6378.48
ViT-gopt-16-SigLIP2-256__webli647564.5178.22
ViT-L-16-SigLIP2-256__webli283023.7778.0
ViT-H-14-378-quickgelu__dfn5b5049108.477.22
ViT-H-14-quickgelu__dfn5b470138.7476.69
nllb-clip-base-siglip__mrl469616.9575.94
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1475.6
ViT-B-16-SigLIP2__webli30385.8175.33
nllb-clip-base-siglip__v1467515.1775.04
ViT-L-16-SigLIP-384__webli339647.672.97
ViT-B-32-SigLIP2-256__webli30613.3172.72
ViT-B-16-SigLIP-i18n-256__webli30296.8772.06
ViT-L-16-SigLIP-256__webli316023.8472.06
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.270.81
ViT-L-14-quickgelu__dfn2b221220.4969.82
ViT-SO400M-14-SigLIP-384__webli441772.1967.54
ViT-B-16-SigLIP-512__webli182826.1766.77
ViT-B-16-SigLIP-384__webli112813.5366.6
ViT-B-16-SigLIP-256__webli11027.1165.67
ViT-B-16-SigLIP__webli10815.7765.29
ViT-H-14__laion2b-s32b-b79k467639.0641.1
ViT-L-14__laion2b-s32b-b82k223320.5634.29
ViT-L-14__laion400m_e31218319.8729.65
ViT-L-14__laion400m_e32221819.7329.56
ViT-B-32__laion2b-s34b-b79k10012.2929.54
ViT-B-32__laion2b_e1610042.3829.36
ViT-B-16-plus-240__laion400m_e3212466.9527.76
ViT-B-16-plus-240__laion400m_e3112636.9427.76
ViT-B-16__laion400m_e329754.9825.67
ViT-B-32__laion400m_e3210032.3525.59
ViT-B-16__laion400m_e319915.0425.53
ViT-B-32__laion400m_e319992.2825.52
ViT-L-14__openai221219.9122.31
RN50x64__openai507948.7922.27
ViT-L-14-336__openai261643.4521.8
RN50x16__openai222115.8720.69
Filipino
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4467.57
nllb-clip-large-siglip__v1422675.0565.64
nllb-clip-base-siglip__mrl469616.9561.21
nllb-clip-base-siglip__v1467515.1759.42
ViT-B-16-SigLIP-i18n-256__webli30296.8736.81
ViT-gopt-16-SigLIP2-384__webli6585146.8435.72
ViT-gopt-16-SigLIP2-256__webli647564.5134.75
ViT-SO400M-14-SigLIP2-378__webli394072.2534.63
ViT-SO400M-16-SigLIP2-512__webli4050107.6734.39
ViT-SO400M-16-SigLIP2-384__webli385456.5734.27
ViT-SO400M-14-SigLIP2__webli362227.6334.14
ViT-SO400M-16-SigLIP2-256__webli361127.8433.98
ViT-L-16-SigLIP2-384__webli305751.730.57
ViT-L-16-SigLIP2-512__webli335892.5930.57
ViT-L-16-SigLIP2-256__webli283023.7730.05
ViT-L-16-SigLIP-384__webli339647.624.92
ViT-L-16-SigLIP-256__webli316023.8424.02
ViT-B-16-SigLIP2__webli30385.8123.37
ViT-B-32-SigLIP2-256__webli30613.3122.69
Finnish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4484.27
nllb-clip-large-siglip__v1422675.0583.93
nllb-clip-base-siglip__mrl469616.9579.41
nllb-clip-base-siglip__v1467515.1778.94
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1475.49
ViT-gopt-16-SigLIP2-384__webli6585146.8463.46
ViT-B-16-SigLIP-i18n-256__webli30296.8763.16
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.263.08
ViT-gopt-16-SigLIP2-256__webli647564.5163.03
ViT-SO400M-16-SigLIP2-384__webli385456.5762.28
ViT-SO400M-16-SigLIP2-256__webli361127.8461.92
ViT-SO400M-14-SigLIP2-378__webli394072.2561.81
ViT-SO400M-14-SigLIP2__webli362227.6361.76
ViT-SO400M-16-SigLIP2-512__webli4050107.6761.05
ViT-L-16-SigLIP2-384__webli305751.757.8
ViT-L-16-SigLIP2-512__webli335892.5957.69
ViT-L-16-SigLIP2-256__webli283023.7757.05
ViT-B-16-SigLIP2__webli30385.8140.26
ViT-B-32-SigLIP2-256__webli30613.3140.06
ViT-L-16-SigLIP-256__webli316023.8431.75
ViT-L-16-SigLIP-384__webli339647.631.74
French
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-384__webli385456.5786.5
ViT-SO400M-16-SigLIP2-512__webli4050107.6786.5
ViT-SO400M-14-SigLIP2-378__webli394072.2586.39
ViT-gopt-16-SigLIP2-384__webli6585146.8486.15
ViT-H-14-378-quickgelu__dfn5b5049108.486.1
nllb-clip-large-siglip__mrl424875.4486.07
nllb-clip-large-siglip__v1422675.0586.06
ViT-H-14-quickgelu__dfn5b470138.7485.89
ViT-L-16-SigLIP2-512__webli335892.5985.67
ViT-SO400M-16-SigLIP2-256__webli361127.8485.67
ViT-gopt-16-SigLIP2-256__webli647564.5185.63
ViT-SO400M-14-SigLIP2__webli362227.6385.39
ViT-L-16-SigLIP2-384__webli305751.785.35
ViT-L-16-SigLIP2-256__webli283023.7784.97
nllb-clip-base-siglip__mrl469616.9583.8
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1482.96
ViT-B-16-SigLIP2__webli30385.8182.91
nllb-clip-base-siglip__v1467515.1782.52
ViT-L-14-quickgelu__dfn2b221220.4981.21
ViT-B-32-SigLIP2-256__webli30613.3180.23
ViT-L-16-SigLIP-384__webli339647.679.85
ViT-B-16-SigLIP-i18n-256__webli30296.8779.47
ViT-L-16-SigLIP-256__webli316023.8479.3
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.277.49
ViT-B-16-SigLIP-512__webli182826.1776.82
ViT-B-16-SigLIP-384__webli112813.5375.94
ViT-B-16-SigLIP__webli10815.7775.3
ViT-B-16-SigLIP-256__webli11027.1175.24
ViT-H-14__laion2b-s32b-b79k467639.0669.33
ViT-SO400M-14-SigLIP-384__webli441772.1964.41
ViT-L-14__laion2b-s32b-b82k223320.5662.86
ViT-L-14__laion400m_e32221819.7359.27
ViT-L-14__laion400m_e31218319.8759.09
ViT-B-16-plus-240__laion400m_e3212466.9558.25
ViT-B-16-plus-240__laion400m_e3112636.9458.25
ViT-B-32__laion2b_e1610042.3856.97
ViT-B-32__laion2b-s34b-b79k10012.2956.21
ViT-B-32__laion400m_e319992.2853.36
ViT-B-16__laion400m_e329754.9853.33
ViT-B-16__laion400m_e319915.0453.26
ViT-B-32__laion400m_e3210032.3553.22
ViT-L-14__openai221219.9146.34
RN50x64__openai507948.7946.3
ViT-L-14-336__openai261643.4545.95
RN50x16__openai222115.8745.69
RN50x4__openai14165.8542.48
RN101__openai11113.2140.16
ViT-B-16__openai9855.0340.1
ViT-B-32__openai10042.2638.27
RN50__openai9132.3937.8
German
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-14-SigLIP2-378__webli394072.2587.32
ViT-SO400M-16-SigLIP2-512__webli4050107.6787.29
ViT-gopt-16-SigLIP2-384__webli6585146.8487.29
ViT-SO400M-16-SigLIP2-384__webli385456.5787.21
ViT-H-14-378-quickgelu__dfn5b5049108.487.18
nllb-clip-large-siglip__mrl424875.4487.14
nllb-clip-large-siglip__v1422675.0587.07
ViT-gopt-16-SigLIP2-256__webli647564.5186.83
ViT-SO400M-14-SigLIP2__webli362227.6386.81
ViT-L-16-SigLIP2-512__webli335892.5986.75
ViT-SO400M-16-SigLIP2-256__webli361127.8486.74
ViT-H-14-quickgelu__dfn5b470138.7486.68
ViT-L-16-SigLIP2-384__webli305751.786.56
ViT-L-16-SigLIP2-256__webli283023.7786.16
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1484.54
nllb-clip-base-siglip__mrl469616.9584.41
ViT-B-16-SigLIP2__webli30385.8184.25
nllb-clip-base-siglip__v1467515.1783.8
ViT-L-14-quickgelu__dfn2b221220.4982.59
ViT-B-32-SigLIP2-256__webli30613.3181.53
ViT-L-16-SigLIP-384__webli339647.681.34
ViT-B-16-SigLIP-i18n-256__webli30296.8781.15
ViT-L-16-SigLIP-256__webli316023.8481.05
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.278.35
ViT-B-16-SigLIP-512__webli182826.1776.56
ViT-B-16-SigLIP-384__webli112813.5376.0
ViT-B-16-SigLIP__webli10815.7775.21
ViT-B-16-SigLIP-256__webli11027.1175.14
ViT-SO400M-14-SigLIP-384__webli441772.1965.86
ViT-H-14__laion2b-s32b-b79k467639.0656.87
ViT-L-14__laion2b-s32b-b82k223320.5647.19
ViT-L-14__laion400m_e32221819.7343.36
ViT-L-14__laion400m_e31218319.8743.0
ViT-B-32__laion2b_e1610042.3841.81
ViT-B-32__laion2b-s34b-b79k10012.2940.43
ViT-B-16-plus-240__laion400m_e3212466.9540.41
ViT-B-16-plus-240__laion400m_e3112636.9440.41
ViT-B-16__laion400m_e319915.0437.71
ViT-B-16__laion400m_e329754.9837.64
ViT-B-32__laion400m_e319992.2836.04
ViT-B-32__laion400m_e3210032.3535.9
RN50x64__openai507948.7934.19
ViT-L-14__openai221219.9133.1
ViT-L-14-336__openai261643.4532.25
RN50x16__openai222115.8730.56
RN50x4__openai14165.8529.2
ViT-B-16__openai9855.0325.77
RN101__openai11113.2125.46
RN50__openai9132.3924.92
ViT-B-32__openai10042.2624.13
Greek
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4474.58
nllb-clip-large-siglip__v1422675.0573.28
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1471.28
nllb-clip-base-siglip__mrl469616.9569.16
nllb-clip-base-siglip__v1467515.1768.21
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.264.69
ViT-gopt-16-SigLIP2-384__webli6585146.8461.64
ViT-gopt-16-SigLIP2-256__webli647564.5161.03
ViT-SO400M-16-SigLIP2-384__webli385456.5760.63
ViT-SO400M-14-SigLIP2-378__webli394072.2560.41
ViT-SO400M-16-SigLIP2-512__webli4050107.6760.1
ViT-SO400M-16-SigLIP2-256__webli361127.8460.06
ViT-SO400M-14-SigLIP2__webli362227.6360.06
ViT-L-16-SigLIP2-384__webli305751.759.44
ViT-L-16-SigLIP2-512__webli335892.5959.44
ViT-L-16-SigLIP2-256__webli283023.7759.43
ViT-B-16-SigLIP-i18n-256__webli30296.8758.78
ViT-B-16-SigLIP2__webli30385.8153.42
ViT-B-32-SigLIP2-256__webli30613.3153.24
Hebrew
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0588.04
nllb-clip-large-siglip__mrl424875.4487.09
nllb-clip-base-siglip__v1467515.1783.93
nllb-clip-base-siglip__mrl469616.9583.84
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1480.78
ViT-B-16-SigLIP-i18n-256__webli30296.8774.59
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.272.73
ViT-SO400M-14-SigLIP2-378__webli394072.2572.25
ViT-gopt-16-SigLIP2-384__webli6585146.8472.19
ViT-SO400M-16-SigLIP2-384__webli385456.5772.15
ViT-SO400M-16-SigLIP2-256__webli361127.8472.08
ViT-SO400M-16-SigLIP2-512__webli4050107.6772.07
ViT-SO400M-14-SigLIP2__webli362227.6372.06
ViT-gopt-16-SigLIP2-256__webli647564.5171.78
ViT-L-16-SigLIP2-512__webli335892.5970.55
ViT-L-16-SigLIP2-384__webli305751.770.03
ViT-L-16-SigLIP2-256__webli283023.7769.34
ViT-B-16-SigLIP2__webli30385.8160.33
ViT-B-32-SigLIP2-256__webli30613.3158.49
Hindi
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4462.02
nllb-clip-large-siglip__v1422675.0561.67
nllb-clip-base-siglip__mrl469616.9558.68
nllb-clip-base-siglip__v1467515.1758.54
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1438.54
ViT-gopt-16-SigLIP2-384__webli6585146.8436.95
ViT-L-16-SigLIP2-512__webli335892.5936.62
ViT-gopt-16-SigLIP2-256__webli647564.5136.06
ViT-L-16-SigLIP2-384__webli305751.735.76
ViT-SO400M-16-SigLIP2-512__webli4050107.6735.34
ViT-SO400M-14-SigLIP2-378__webli394072.2535.17
ViT-SO400M-16-SigLIP2-384__webli385456.5734.94
ViT-L-16-SigLIP2-256__webli283023.7734.91
ViT-SO400M-16-SigLIP2-256__webli361127.8434.19
ViT-SO400M-14-SigLIP2__webli362227.6333.56
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.232.06
ViT-B-16-SigLIP-i18n-256__webli30296.8731.85
ViT-B-16-SigLIP2__webli30385.8127.87
ViT-B-32-SigLIP2-256__webli30613.3127.08
Hungarian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4485.59
nllb-clip-large-siglip__v1422675.0585.25
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1481.74
nllb-clip-base-siglip__mrl469616.9580.34
nllb-clip-base-siglip__v1467515.1780.14
ViT-gopt-16-SigLIP2-384__webli6585146.8474.94
ViT-SO400M-14-SigLIP2-378__webli394072.2574.2
ViT-gopt-16-SigLIP2-256__webli647564.5174.03
ViT-SO400M-16-SigLIP2-512__webli4050107.6773.96
ViT-B-16-SigLIP-i18n-256__webli30296.8773.95
ViT-SO400M-16-SigLIP2-384__webli385456.5773.9
ViT-SO400M-16-SigLIP2-256__webli361127.8473.59
ViT-SO400M-14-SigLIP2__webli362227.6373.12
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.272.5
ViT-L-16-SigLIP2-512__webli335892.5972.33
ViT-L-16-SigLIP2-384__webli305751.771.83
ViT-L-16-SigLIP2-256__webli283023.7770.57
ViT-B-16-SigLIP2__webli30385.8158.31
ViT-B-32-SigLIP2-256__webli30613.3156.74
ViT-L-16-SigLIP-384__webli339647.638.26
ViT-L-16-SigLIP-256__webli316023.8437.97
ViT-H-14-quickgelu__dfn5b470138.7428.75
ViT-H-14-378-quickgelu__dfn5b5049108.428.26
ViT-B-16-SigLIP-512__webli182826.1724.88
ViT-B-16-SigLIP-384__webli112813.5324.39
ViT-B-16-SigLIP__webli10815.7724.29
ViT-B-16-SigLIP-256__webli11027.1124.16
Indonesian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0585.46
ViT-SO400M-14-SigLIP2-378__webli394072.2585.12
nllb-clip-large-siglip__mrl424875.4485.01
ViT-SO400M-16-SigLIP2-384__webli385456.5784.99
ViT-SO400M-16-SigLIP2-512__webli4050107.6784.65
ViT-gopt-16-SigLIP2-384__webli6585146.8484.62
ViT-L-16-SigLIP2-384__webli305751.784.58
ViT-L-16-SigLIP2-512__webli335892.5984.11
ViT-gopt-16-SigLIP2-256__webli647564.5184.1
ViT-SO400M-16-SigLIP2-256__webli361127.8484.06
ViT-L-16-SigLIP2-256__webli283023.7783.69
ViT-SO400M-14-SigLIP2__webli362227.6383.61
nllb-clip-base-siglip__v1467515.1782.31
nllb-clip-base-siglip__mrl469616.9581.97
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1480.93
ViT-B-16-SigLIP2__webli30385.8179.84
ViT-B-16-SigLIP-i18n-256__webli30296.8777.12
ViT-B-32-SigLIP2-256__webli30613.3177.02
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.274.15
ViT-L-16-SigLIP-384__webli339647.671.44
ViT-L-16-SigLIP-256__webli316023.8469.94
ViT-H-14-378-quickgelu__dfn5b5049108.465.87
ViT-H-14-quickgelu__dfn5b470138.7465.19
ViT-B-16-SigLIP-512__webli182826.1759.95
ViT-B-16-SigLIP-384__webli112813.5359.38
ViT-B-16-SigLIP-256__webli11027.1157.88
ViT-B-16-SigLIP__webli10815.7757.52
ViT-SO400M-14-SigLIP-384__webli441772.1954.11
ViT-L-14-quickgelu__dfn2b221220.4950.02
ViT-H-14__laion2b-s32b-b79k467639.0623.25
Italian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-512__webli4050107.6787.17
ViT-SO400M-14-SigLIP2-378__webli394072.2586.91
ViT-gopt-16-SigLIP2-384__webli6585146.8486.83
ViT-SO400M-16-SigLIP2-384__webli385456.5786.77
ViT-L-16-SigLIP2-512__webli335892.5986.67
ViT-gopt-16-SigLIP2-256__webli647564.5186.42
ViT-L-16-SigLIP2-384__webli305751.786.35
ViT-H-14-378-quickgelu__dfn5b5049108.486.34
ViT-SO400M-16-SigLIP2-256__webli361127.8486.18
nllb-clip-large-siglip__v1422675.0586.17
ViT-SO400M-14-SigLIP2__webli362227.6385.84
nllb-clip-large-siglip__mrl424875.4485.8
ViT-L-16-SigLIP2-256__webli283023.7785.7
ViT-H-14-quickgelu__dfn5b470138.7485.67
ViT-B-16-SigLIP2__webli30385.8183.32
nllb-clip-base-siglip__mrl469616.9582.95
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1482.73
nllb-clip-base-siglip__v1467515.1782.72
ViT-L-16-SigLIP-384__webli339647.681.07
ViT-B-32-SigLIP2-256__webli30613.3180.8
ViT-L-14-quickgelu__dfn2b221220.4980.6
ViT-L-16-SigLIP-256__webli316023.8480.35
ViT-B-16-SigLIP-i18n-256__webli30296.8778.79
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.276.62
ViT-B-16-SigLIP-512__webli182826.1776.51
ViT-B-16-SigLIP-384__webli112813.5376.08
ViT-B-16-SigLIP__webli10815.7775.29
ViT-B-16-SigLIP-256__webli11027.1175.29
ViT-SO400M-14-SigLIP-384__webli441772.1974.84
ViT-H-14__laion2b-s32b-b79k467639.0656.32
ViT-L-14__laion2b-s32b-b82k223320.5647.25
ViT-L-14__laion400m_e32221819.7343.09
ViT-L-14__laion400m_e31218319.8742.99
ViT-B-16-plus-240__laion400m_e3212466.9540.29
ViT-B-16-plus-240__laion400m_e3112636.9440.29
ViT-B-32__laion2b_e1610042.3839.67
ViT-B-32__laion2b-s34b-b79k10012.2939.03
ViT-B-16__laion400m_e329754.9836.14
ViT-B-16__laion400m_e319915.0435.89
ViT-B-32__laion400m_e3210032.3535.59
ViT-B-32__laion400m_e319992.2835.56
RN50x64__openai507948.7933.53
ViT-L-14__openai221219.9132.19
ViT-L-14-336__openai261643.4530.95
RN50x16__openai222115.8728.85
RN50x4__openai14165.8525.75
ViT-B-16__openai9855.0325.18
RN101__openai11113.2124.48
RN50__openai9132.3923.89
ViT-B-32__openai10042.2623.39
Japanese
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1483.95
nllb-clip-large-siglip__v1422675.0582.21
nllb-clip-large-siglip__mrl424875.4481.55
nllb-clip-base-siglip__v1467515.1778.72
nllb-clip-base-siglip__mrl469616.9578.53
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.275.93
ViT-gopt-16-SigLIP2-384__webli6585146.8466.86
ViT-SO400M-16-SigLIP2-384__webli385456.5765.59
ViT-SO400M-16-SigLIP2-512__webli4050107.6765.48
ViT-SO400M-14-SigLIP2-378__webli394072.2565.36
ViT-gopt-16-SigLIP2-256__webli647564.5164.47
ViT-SO400M-16-SigLIP2-256__webli361127.8464.17
ViT-L-16-SigLIP2-384__webli305751.764.08
ViT-L-16-SigLIP2-256__webli283023.7763.69
ViT-L-16-SigLIP2-512__webli335892.5963.33
ViT-SO400M-14-SigLIP2__webli362227.6363.02
ViT-B-16-SigLIP-i18n-256__webli30296.8758.39
ViT-B-16-SigLIP2__webli30385.8156.38
ViT-B-32-SigLIP2-256__webli30613.3153.16
Korean
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4480.56
nllb-clip-large-siglip__v1422675.0580.53
nllb-clip-base-siglip__mrl469616.9577.09
ViT-SO400M-14-SigLIP2-378__webli394072.2577.08
ViT-SO400M-16-SigLIP2-512__webli4050107.6776.97
ViT-SO400M-16-SigLIP2-384__webli385456.5776.92
nllb-clip-base-siglip__v1467515.1776.58
ViT-SO400M-16-SigLIP2-256__webli361127.8476.2
ViT-SO400M-14-SigLIP2__webli362227.6375.95
ViT-L-16-SigLIP2-512__webli335892.5975.86
ViT-L-16-SigLIP2-384__webli305751.775.67
ViT-gopt-16-SigLIP2-384__webli6585146.8475.49
ViT-gopt-16-SigLIP2-256__webli647564.5174.6
ViT-L-16-SigLIP2-256__webli283023.7774.52
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1473.88
ViT-B-16-SigLIP2__webli30385.8171.09
ViT-B-16-SigLIP-i18n-256__webli30296.8768.87
ViT-B-32-SigLIP2-256__webli30613.3167.94
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.266.39
Maori
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4448.43
nllb-clip-large-siglip__v1422675.0546.12
nllb-clip-base-siglip__mrl469616.9542.8
nllb-clip-base-siglip__v1467515.1740.85
Norwegian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4481.36
nllb-clip-large-siglip__v1422675.0580.96
nllb-clip-base-siglip__mrl469616.9577.65
nllb-clip-base-siglip__v1467515.1776.39
ViT-gopt-16-SigLIP2-384__webli6585146.8475.97
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1475.44
ViT-gopt-16-SigLIP2-256__webli647564.5175.31
ViT-SO400M-16-SigLIP2-384__webli385456.5775.0
ViT-SO400M-16-SigLIP2-512__webli4050107.6774.96
ViT-SO400M-14-SigLIP2-378__webli394072.2574.92
ViT-SO400M-16-SigLIP2-256__webli361127.8474.44
ViT-SO400M-14-SigLIP2__webli362227.6374.37
ViT-L-16-SigLIP2-512__webli335892.5973.11
ViT-L-16-SigLIP2-384__webli305751.772.63
ViT-L-16-SigLIP2-256__webli283023.7771.71
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.267.81
ViT-B-16-SigLIP-i18n-256__webli30296.8765.55
ViT-B-16-SigLIP2__webli30385.8162.56
ViT-B-32-SigLIP2-256__webli30613.3160.94
ViT-H-14-quickgelu__dfn5b470138.7459.62
ViT-H-14-378-quickgelu__dfn5b5049108.459.49
ViT-L-16-SigLIP-256__webli316023.8446.3
ViT-L-16-SigLIP-384__webli339647.645.75
ViT-L-14-quickgelu__dfn2b221220.4942.55
ViT-B-16-SigLIP-512__webli182826.1735.33
ViT-B-16-SigLIP__webli10815.7735.01
ViT-B-16-SigLIP-384__webli112813.5334.94
ViT-B-16-SigLIP-256__webli11027.1134.39
Persian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4479.52
nllb-clip-large-siglip__v1422675.0578.99
ViT-SO400M-16-SigLIP2-512__webli4050107.6776.32
ViT-SO400M-16-SigLIP2-384__webli385456.5776.3
ViT-SO400M-14-SigLIP2-378__webli394072.2576.11
ViT-L-16-SigLIP2-512__webli335892.5975.56
nllb-clip-base-siglip__mrl469616.9575.38
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1474.92
nllb-clip-base-siglip__v1467515.1774.86
ViT-L-16-SigLIP2-384__webli305751.774.73
ViT-SO400M-16-SigLIP2-256__webli361127.8474.32
ViT-gopt-16-SigLIP2-384__webli6585146.8474.31
ViT-SO400M-14-SigLIP2__webli362227.6373.42
ViT-gopt-16-SigLIP2-256__webli647564.5172.56
ViT-L-16-SigLIP2-256__webli283023.7771.9
ViT-B-16-SigLIP-i18n-256__webli30296.8769.79
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.268.55
ViT-B-16-SigLIP2__webli30385.8168.26
ViT-B-32-SigLIP2-256__webli30613.3165.16
Polish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4483.49
ViT-gopt-16-SigLIP2-384__webli6585146.8483.45
nllb-clip-large-siglip__v1422675.0583.11
ViT-SO400M-16-SigLIP2-384__webli385456.5782.99
ViT-SO400M-16-SigLIP2-512__webli4050107.6782.96
ViT-SO400M-14-SigLIP2-378__webli394072.2582.93
ViT-gopt-16-SigLIP2-256__webli647564.5182.61
ViT-L-16-SigLIP2-512__webli335892.5982.26
ViT-SO400M-16-SigLIP2-256__webli361127.8482.24
ViT-L-16-SigLIP2-384__webli305751.782.03
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1482.03
ViT-SO400M-14-SigLIP2__webli362227.6381.92
ViT-L-16-SigLIP2-256__webli283023.7781.27
nllb-clip-base-siglip__mrl469616.9580.0
nllb-clip-base-siglip__v1467515.1779.65
ViT-B-16-SigLIP-i18n-256__webli30296.8776.75
ViT-B-16-SigLIP2__webli30385.8176.52
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.275.1
ViT-B-32-SigLIP2-256__webli30613.3173.9
ViT-H-14-378-quickgelu__dfn5b5049108.465.03
ViT-H-14-quickgelu__dfn5b470138.7464.89
ViT-L-16-SigLIP-256__webli316023.8451.6
ViT-L-16-SigLIP-384__webli339647.651.29
ViT-L-14-quickgelu__dfn2b221220.4946.15
ViT-B-16-SigLIP-512__webli182826.1741.55
ViT-B-16-SigLIP-384__webli112813.5341.17
ViT-B-16-SigLIP-256__webli11027.1140.9
ViT-B-16-SigLIP__webli10815.7740.76
Portuguese
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-14-SigLIP2-378__webli394072.2582.12
ViT-SO400M-16-SigLIP2-512__webli4050107.6781.84
ViT-L-16-SigLIP2-512__webli335892.5981.69
ViT-SO400M-16-SigLIP2-384__webli385456.5781.69
ViT-gopt-16-SigLIP2-384__webli6585146.8481.54
ViT-L-16-SigLIP2-384__webli305751.781.39
ViT-SO400M-16-SigLIP2-256__webli361127.8480.56
ViT-gopt-16-SigLIP2-256__webli647564.5180.34
ViT-L-16-SigLIP2-256__webli283023.7780.02
nllb-clip-large-siglip__mrl424875.4479.99
ViT-SO400M-14-SigLIP2__webli362227.6379.93
ViT-H-14-378-quickgelu__dfn5b5049108.479.61
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1479.12
ViT-H-14-quickgelu__dfn5b470138.7478.87
nllb-clip-large-siglip__v1422675.0578.85
ViT-B-16-SigLIP2__webli30385.8177.54
ViT-B-16-SigLIP-i18n-256__webli30296.8775.31
nllb-clip-base-siglip__mrl469616.9575.26
ViT-B-32-SigLIP2-256__webli30613.3174.82
ViT-L-16-SigLIP-384__webli339647.674.48
nllb-clip-base-siglip__v1467515.1774.47
ViT-L-14-quickgelu__dfn2b221220.4973.92
ViT-L-16-SigLIP-256__webli316023.8473.58
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.273.02
ViT-B-16-SigLIP-512__webli182826.1771.44
ViT-B-16-SigLIP-384__webli112813.5371.16
ViT-B-16-SigLIP-256__webli11027.1169.69
ViT-B-16-SigLIP__webli10815.7769.32
ViT-SO400M-14-SigLIP-384__webli441772.1959.86
ViT-H-14__laion2b-s32b-b79k467639.0645.49
ViT-L-14__laion2b-s32b-b82k223320.5637.86
ViT-L-14__laion400m_e32221819.7336.01
ViT-L-14__laion400m_e31218319.8735.75
ViT-B-16-plus-240__laion400m_e3212466.9533.25
ViT-B-16-plus-240__laion400m_e3112636.9433.25
ViT-B-32__laion2b_e1610042.3832.83
ViT-B-32__laion2b-s34b-b79k10012.2932.62
ViT-B-32__laion400m_e3210032.3530.86
ViT-B-32__laion400m_e319992.2830.8
RN50x64__openai507948.7930.58
ViT-B-16__laion400m_e329754.9830.18
ViT-B-16__laion400m_e319915.0429.93
ViT-L-14__openai221219.9128.88
ViT-L-14-336__openai261643.4528.49
RN50x16__openai222115.8723.9
RN50x4__openai14165.8522.94
ViT-B-16__openai9855.0322.55
RN50__openai9132.3921.85
ViT-B-32__openai10042.2621.3
RN101__openai11113.2121.14
Romanian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0589.38
nllb-clip-large-siglip__mrl424875.4488.86
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1485.37
nllb-clip-base-siglip__v1467515.1784.92
nllb-clip-base-siglip__mrl469616.9584.49
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.277.92
ViT-gopt-16-SigLIP2-384__webli6585146.8474.98
ViT-gopt-16-SigLIP2-256__webli647564.5174.33
ViT-SO400M-14-SigLIP2-378__webli394072.2574.05
ViT-SO400M-16-SigLIP2-512__webli4050107.6774.03
ViT-SO400M-16-SigLIP2-384__webli385456.5773.94
ViT-SO400M-14-SigLIP2__webli362227.6373.27
ViT-SO400M-16-SigLIP2-256__webli361127.8473.22
ViT-L-16-SigLIP2-512__webli335892.5972.91
ViT-L-16-SigLIP2-384__webli305751.772.43
ViT-L-16-SigLIP2-256__webli283023.7771.93
ViT-B-16-SigLIP-i18n-256__webli30296.8771.5
ViT-B-16-SigLIP2__webli30385.8158.28
ViT-B-32-SigLIP2-256__webli30613.3156.54
ViT-H-14-378-quickgelu__dfn5b5049108.456.12
ViT-H-14-quickgelu__dfn5b470138.7455.53
ViT-L-14-quickgelu__dfn2b221220.4934.96
ViT-L-16-SigLIP-384__webli339647.626.33
ViT-L-16-SigLIP-256__webli316023.8426.05
ViT-B-16-SigLIP-256__webli11027.1121.32
ViT-B-16-SigLIP-512__webli182826.1721.04
ViT-B-16-SigLIP-384__webli112813.5320.76
ViT-B-16-SigLIP__webli10815.7720.56
Russian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-384__webli385456.5784.54
ViT-SO400M-14-SigLIP2-378__webli394072.2584.41
ViT-SO400M-16-SigLIP2-512__webli4050107.6784.36
ViT-gopt-16-SigLIP2-384__webli6585146.8484.31
ViT-L-16-SigLIP2-512__webli335892.5984.22
ViT-SO400M-16-SigLIP2-256__webli361127.8483.9
ViT-L-16-SigLIP2-384__webli305751.783.69
ViT-SO400M-14-SigLIP2__webli362227.6383.5
nllb-clip-large-siglip__mrl424875.4483.31
ViT-gopt-16-SigLIP2-256__webli647564.5183.21
ViT-L-16-SigLIP2-256__webli283023.7783.11
nllb-clip-large-siglip__v1422675.0582.7
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1482.69
ViT-B-16-SigLIP2__webli30385.8180.91
nllb-clip-base-siglip__mrl469616.9579.75
ViT-B-16-SigLIP-i18n-256__webli30296.8779.35
nllb-clip-base-siglip__v1467515.1778.91
ViT-B-32-SigLIP2-256__webli30613.3178.06
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.276.44
ViT-H-14-378-quickgelu__dfn5b5049108.442.81
ViT-H-14-quickgelu__dfn5b470138.7442.1
ViT-L-16-SigLIP-256__webli316023.8424.95
ViT-L-16-SigLIP-384__webli339647.624.25
ViT-B-16-SigLIP-256__webli11027.1120.85
ViT-B-16-SigLIP__webli10815.7720.44
ViT-B-16-SigLIP-512__webli182826.1720.41
Spanish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-14-SigLIP2-378__webli394072.2585.47
ViT-SO400M-16-SigLIP2-384__webli385456.5785.44
ViT-L-16-SigLIP2-512__webli335892.5985.32
ViT-SO400M-16-SigLIP2-512__webli4050107.6785.22
ViT-gopt-16-SigLIP2-384__webli6585146.8485.15
ViT-L-16-SigLIP2-384__webli305751.784.81
ViT-gopt-16-SigLIP2-256__webli647564.5184.68
ViT-SO400M-16-SigLIP2-256__webli361127.8484.6
ViT-SO400M-14-SigLIP2__webli362227.6384.55
ViT-H-14-378-quickgelu__dfn5b5049108.484.27
ViT-L-16-SigLIP2-256__webli283023.7784.15
ViT-H-14-quickgelu__dfn5b470138.7483.87
nllb-clip-large-siglip__mrl424875.4483.74
ViT-B-16-SigLIP2__webli30385.8183.61
nllb-clip-large-siglip__v1422675.0583.15
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1481.7
nllb-clip-base-siglip__mrl469616.9580.91
ViT-B-32-SigLIP2-256__webli30613.3180.73
ViT-L-16-SigLIP-384__webli339647.680.69
ViT-L-16-SigLIP-256__webli316023.8480.3
nllb-clip-base-siglip__v1467515.1779.8
ViT-B-16-SigLIP-i18n-256__webli30296.8779.71
ViT-L-14-quickgelu__dfn2b221220.4979.64
ViT-B-16-SigLIP-384__webli112813.5378.0
ViT-B-16-SigLIP-512__webli182826.1777.83
ViT-B-16-SigLIP__webli10815.7776.87
ViT-B-16-SigLIP-256__webli11027.1176.66
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.275.99
ViT-SO400M-14-SigLIP-384__webli441772.1971.96
ViT-H-14__laion2b-s32b-b79k467639.0662.06
ViT-L-14__laion2b-s32b-b82k223320.5653.78
ViT-L-14__laion400m_e32221819.7350.13
ViT-L-14__laion400m_e31218319.8750.0
ViT-B-16-plus-240__laion400m_e3212466.9547.39
ViT-B-16-plus-240__laion400m_e3112636.9447.39
ViT-B-32__laion2b_e1610042.3846.47
ViT-B-32__laion2b-s34b-b79k10012.2945.68
ViT-B-16__laion400m_e319915.0444.0
ViT-B-16__laion400m_e329754.9843.98
ViT-B-32__laion400m_e3210032.3543.8
ViT-B-32__laion400m_e319992.2843.73
RN50x64__openai507948.7943.01
ViT-L-14__openai221219.9142.96
ViT-L-14-336__openai261643.4541.67
RN50x16__openai222115.8740.21
RN50x4__openai14165.8536.06
ViT-B-16__openai9855.0335.67
RN101__openai11113.2134.62
ViT-B-32__openai10042.2632.6
RN50__openai9132.3931.79
Swahili
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4469.51
nllb-clip-large-siglip__v1422675.0568.44
nllb-clip-base-siglip__mrl469616.9566.09
nllb-clip-base-siglip__v1467515.1763.98
ViT-B-16-SigLIP-i18n-256__webli30296.8721.64
Swedish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4477.12
nllb-clip-large-siglip__v1422675.0576.37
nllb-clip-base-siglip__mrl469616.9573.41
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1472.83
ViT-gopt-16-SigLIP2-384__webli6585146.8472.51
ViT-gopt-16-SigLIP2-256__webli647564.5172.2
ViT-SO400M-14-SigLIP2-378__webli394072.2572.1
ViT-SO400M-16-SigLIP2-384__webli385456.5772.06
ViT-L-16-SigLIP2-512__webli335892.5971.84
ViT-L-16-SigLIP2-384__webli305751.771.7
ViT-SO400M-16-SigLIP2-256__webli361127.8471.7
ViT-SO400M-16-SigLIP2-512__webli4050107.6771.61
nllb-clip-base-siglip__v1467515.1771.51
ViT-SO400M-14-SigLIP2__webli362227.6371.45
ViT-L-16-SigLIP2-256__webli283023.7771.23
ViT-B-16-SigLIP-i18n-256__webli30296.8767.48
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.266.93
ViT-B-16-SigLIP2__webli30385.8166.37
ViT-B-32-SigLIP2-256__webli30613.3164.86
ViT-H-14-378-quickgelu__dfn5b5049108.462.35
ViT-H-14-quickgelu__dfn5b470138.7461.51
ViT-L-16-SigLIP-256__webli316023.8456.74
ViT-L-16-SigLIP-384__webli339647.655.92
ViT-B-16-SigLIP-512__webli182826.1748.5
ViT-B-16-SigLIP__webli10815.7748.38
ViT-B-16-SigLIP-256__webli11027.1148.06
ViT-B-16-SigLIP-384__webli112813.5347.99
ViT-L-14-quickgelu__dfn2b221220.4947.93
ViT-SO400M-14-SigLIP-384__webli441772.1929.98
Telugu
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4464.32
nllb-clip-large-siglip__v1422675.0562.34
nllb-clip-base-siglip__mrl469616.9560.72
nllb-clip-base-siglip__v1467515.1758.8
Thai
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4479.99
nllb-clip-large-siglip__v1422675.0579.07
nllb-clip-base-siglip__mrl469616.9576.13
nllb-clip-base-siglip__v1467515.1775.23
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1474.04
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.266.03
ViT-SO400M-14-SigLIP2-378__webli394072.2545.87
ViT-L-16-SigLIP2-384__webli305751.745.69
ViT-SO400M-16-SigLIP2-384__webli385456.5745.52
ViT-SO400M-16-SigLIP2-512__webli4050107.6744.96
ViT-L-16-SigLIP2-512__webli335892.5944.75
ViT-SO400M-16-SigLIP2-256__webli361127.8444.66
ViT-SO400M-14-SigLIP2__webli362227.6343.99
ViT-L-16-SigLIP2-256__webli283023.7743.91
ViT-gopt-16-SigLIP2-384__webli6585146.8443.06
ViT-gopt-16-SigLIP2-256__webli647564.5141.86
ViT-B-16-SigLIP-i18n-256__webli30296.8741.1
ViT-B-16-SigLIP2__webli30385.8137.35
ViT-B-32-SigLIP2-256__webli30613.3135.28
Turkish
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__mrl424875.4483.91
nllb-clip-large-siglip__v1422675.0583.74
nllb-clip-base-siglip__mrl469616.9581.26
nllb-clip-base-siglip__v1467515.1780.21
ViT-SO400M-16-SigLIP2-512__webli4050107.6779.34
ViT-SO400M-14-SigLIP2-378__webli394072.2579.22
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1478.9
ViT-SO400M-16-SigLIP2-384__webli385456.5778.85
ViT-SO400M-16-SigLIP2-256__webli361127.8478.29
ViT-gopt-16-SigLIP2-384__webli6585146.8478.27
ViT-gopt-16-SigLIP2-256__webli647564.5178.0
ViT-SO400M-14-SigLIP2__webli362227.6377.81
ViT-L-16-SigLIP2-512__webli335892.5977.67
ViT-L-16-SigLIP2-384__webli305751.777.33
ViT-L-16-SigLIP2-256__webli283023.7776.42
ViT-B-16-SigLIP-i18n-256__webli30296.8772.44
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.269.84
ViT-B-16-SigLIP2__webli30385.8169.83
ViT-B-32-SigLIP2-256__webli30613.3167.13
ViT-H-14-378-quickgelu__dfn5b5049108.444.43
ViT-H-14-quickgelu__dfn5b470138.7443.87
ViT-L-16-SigLIP-384__webli339647.635.1
ViT-L-16-SigLIP-256__webli316023.8434.92
ViT-L-14-quickgelu__dfn2b221220.4925.2
ViT-B-16-SigLIP-512__webli182826.1724.55
ViT-B-16-SigLIP__webli10815.7724.13
ViT-B-16-SigLIP-384__webli112813.5324.08
ViT-B-16-SigLIP-256__webli11027.1123.95
Ukrainian
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
nllb-clip-large-siglip__v1422675.0583.92
nllb-clip-large-siglip__mrl424875.4483.88
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1483.2
nllb-clip-base-siglip__mrl469616.9579.99
nllb-clip-base-siglip__v1467515.1779.31
ViT-SO400M-14-SigLIP2-378__webli394072.2578.73
ViT-SO400M-16-SigLIP2-384__webli385456.5778.33
ViT-SO400M-16-SigLIP2-512__webli4050107.6777.95
ViT-SO400M-16-SigLIP2-256__webli361127.8477.56
ViT-SO400M-14-SigLIP2__webli362227.6377.49
ViT-gopt-16-SigLIP2-384__webli6585146.8477.02
ViT-gopt-16-SigLIP2-256__webli647564.5176.87
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.276.31
ViT-L-16-SigLIP2-512__webli335892.5975.91
ViT-L-16-SigLIP2-384__webli305751.775.75
ViT-L-16-SigLIP2-256__webli283023.7775.1
ViT-B-16-SigLIP-i18n-256__webli30296.8773.3
ViT-B-16-SigLIP2__webli30385.8165.28
ViT-B-32-SigLIP2-256__webli30613.3163.95
Vietnamese
ModelMemory (MiB)Execution Time (ms)Recall (%)Pareto Optimal
ViT-SO400M-16-SigLIP2-384__webli385456.5785.86
ViT-SO400M-14-SigLIP2-378__webli394072.2585.73
ViT-SO400M-16-SigLIP2-512__webli4050107.6785.67
ViT-gopt-16-SigLIP2-384__webli6585146.8485.5
ViT-L-16-SigLIP2-384__webli305751.784.93
ViT-SO400M-16-SigLIP2-256__webli361127.8484.84
ViT-L-16-SigLIP2-512__webli335892.5984.78
ViT-SO400M-14-SigLIP2__webli362227.6384.34
ViT-gopt-16-SigLIP2-256__webli647564.5184.33
ViT-L-16-SigLIP2-256__webli283023.7783.93
nllb-clip-large-siglip__mrl424875.4483.69
nllb-clip-large-siglip__v1422675.0583.19
XLM-Roberta-Large-ViT-H-14__frozen_laion5b_s13b_b90k401439.1481.88
ViT-B-16-SigLIP2__webli30385.8180.88
nllb-clip-base-siglip__mrl469616.9579.79
nllb-clip-base-siglip__v1467515.1779.38
ViT-B-32-SigLIP2-256__webli30613.3177.73
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k30303.275.18
ViT-B-16-SigLIP-i18n-256__webli30296.8773.05
note

Feel free to make a feature request if there's a model you want to use that we don't currently support.