SAVRN
Search Contact SAVRN

Open-weight model · Image feature extraction

clip-vitb-mini-distilled

by AbstractPhila AbstractPhil/clip-vitb-mini-distilled

clip-vitb-mini-distilled is an open-weight model for image feature extraction from AbstractPhila, released under Apache License 2.0. It has 9M parameters. At 16-bit it needs about 0 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 64 downloads a month.

An 8.66M-parameter ViT image encoder (10.0% of a CLIP-B/16 image tower) producing 512-d embeddings compatible with the text tower.

Parameters9M
Context—
Weights2.9 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads64

Runs On

What it takes to serve clip-vitb-mini-distilled (9M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 0.0 GB 0.0 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 0.0 GB 0.0 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.0 GB 0.0 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 9, 2026.

clip-vitb-mini-distilled on every accelerator the SAVRN Index prices, at every precision

Model Card

By AbstractPhila, published under apache-2.0, revision 791d0e36cdf7.

An 8.66M-parameter ViT image encoder (10.0% of a CLIP-B/16 image tower) producing 512-d embeddings compatible with the CLIP-B/16 LAION-2B text tower. The primary checkpoint was distilled on CC12M (10,968,539 images) against the generalized-Procrustes consensus of five CLIP teachers — never against the deployment teacher — and carries a frozen 512×512 rotation that maps its outputs into the deployment frame, where it outperforms the student distilled directly against that teacher on every task gauge, both seeds (full tables below).

Quick start (AutoModel)

import torch
from transformers import AutoModel, AutoImageProcessor

repo = "AbstractPhil/clip-vitb-mini-distilled"
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()
proc = AutoImageProcessor.from_pretrained(repo)

px = proc(images=[img], return_tensors="pt")["pixel_values"]
emb = model.get_image_features(px)     # (1, 512), L2-normalized,
                                       # deployment (LAION-B/16) frame

The rotation is applied by default (config.apply_rotation); pass apply_rotation=False to get_image_features for the raw consensus-frame embedding. Weights are safetensors; the modeling code is in this repo (modeling_clip_mini.py, configuration_clip_mini.py).

Read the full model card (1,696 words)

Configuration

Architecture
ClipMiniModel
Layers
12
Hidden size
240
Attention heads
4
Model type
clip_vitb_mini

Identity and Version

Repository
AbstractPhil/clip-vitb-mini-distilled
Publisher
AbstractPhila
Task
Image feature extraction
Modality
Other
Library
transformers
Parameters
9M parameters
Languages
vit
Revision
791d0e36cdf7e81d2e41128d5d0635636cb6ad08
First published
2026-07-27
Last updated
2026-10-09

Files and Weights

163 files, 2.9 GB in total. The weights are 85 files totalling 2.9 GB in pt, safetensors.

Weights85 files · 2.9 GB
Configuration25 files · 324.2 KB
Documentation3 files · 49.4 KB
Other49 files · 128.2 KB
Repository1 file · 1.5 KB
Every file
FileTypeSizeSHA-256
cc12m/affinity_kl_s0_t88000.ptWeights34.7 MB ca8acd50734c
cc12m/affinity_kl_s1_t88000.ptWeights34.7 MB 05e7fa4ca78b
cc12m/cbert_full_s0_t88000.ptWeights34.7 MB 77fcb3ecfde2
cc12m/cbert_full_s1_t88000.ptWeights34.7 MB a9537d858d00
cc12m/consensus_gpa_s0_t88000.ptWeights34.7 MB d2e91dc09561
cc12m/consensus_gpa_s1_t88000.ptWeights34.7 MB 6360fc7b4b01
cc12m/consensus_nce_mse_s0_t88000.ptWeights34.7 MB 3ccc6205e54e
cc12m/consensus_nce_mse_s1_t88000.ptWeights34.7 MB 9eb00655d420
cc12m/feature_mse_s0_t88000.ptWeights34.7 MB f79532ccb186
cc12m/feature_mse_s1_t88000.ptWeights34.7 MB f98365e62210
cc12m/infonce_s0_t88000.ptWeights34.7 MB 2c77821b7bd6
cc12m/infonce_s1_t88000.ptWeights34.7 MB 0e31e0386ca6
cc12m/mid/affinity_kl_s1_mid.ptWeights34.7 MB dcf3510378f0
cc12m/mid/cbert_full_s0_mid.ptWeights34.7 MB 3e301e410b3f
cc12m/mid/cbert_full_s1_mid.ptWeights34.7 MB e5aa2c470b79
cc12m/mid/consensus_gpa_s0_mid.ptWeights34.7 MB 7e88c0489dee
cc12m/mid/consensus_gpa_s1_mid.ptWeights34.7 MB 2a03b06f15ba
cc12m/mid/consensus_nce_mse_s0_mid.ptWeights34.7 MB 619d4cc11079
cc12m/mid/consensus_nce_mse_s1_mid.ptWeights34.7 MB efd166647cf0
cc12m/mid/feature_mse_s1_mid.ptWeights34.7 MB 4e3b8872ef4e
cc12m/mid/infonce_s1_mid.ptWeights34.7 MB add4851d568a
cc12m/mid/nce_mse_cv_s0_mid.ptWeights34.7 MB 37d58539e9f1
cc12m/mid/nce_mse_cv_s1_mid.ptWeights34.7 MB d43fe808a8b1
cc12m/mid/siglip_pairwise_s0_mid.ptWeights35.0 MB d3027647873b
cc12m/mid/siglip_pairwise_s1_mid.ptWeights35.0 MB 4b41dddf74b5
cc12m/mid/x3_autograd_s0_mid.ptWeights34.7 MB 10e2eace6a7d
cc12m/mid/x3_autograd_s1_mid.ptWeights34.7 MB 3c49a3dfde29
cc12m/nce_mse_cv_s0_t88000.ptWeights34.7 MB 8e4e443c41ec
cc12m/nce_mse_cv_s1_t88000.ptWeights34.7 MB e8c8935f40cb
cc12m/rotation_s0.ptWeights1.0 MB a5279eacde1b
cc12m/rotation_s1.ptWeights1.0 MB 8bc1e2a70e04
cc12m/siglip_pairwise_s0_t88000.ptWeights35.0 MB a1d9200a5118
cc12m/siglip_pairwise_s1_t88000.ptWeights35.0 MB 90435a53f5ee
cc12m/x3_autograd_s0_t88000.ptWeights34.7 MB 713c9c43ff49
cc12m/x3_autograd_s1_t88000.ptWeights34.7 MB 3daa6a16bfd2
coco/affinity_kl_s0_t8000.ptWeights34.7 MB ff0f56baf44f
coco/affinity_kl_s1_t8000.ptWeights34.7 MB 4a4e9e703d21
coco/blueprint_s0_t8000.ptWeights34.7 MB 91e1cb3887da
coco/blueprint_s1_t8000.ptWeights34.7 MB d46d48f8f22e
coco/cbert_full_s0_t8000.ptWeights34.7 MB c915216eb4c1
coco/cbert_full_s1_t8000.ptWeights34.7 MB 9523461c7e96
coco/consensus_gpa_s0_t8000.ptWeights34.7 MB 74d699248455
coco/consensus_gpa_s1_t8000.ptWeights34.7 MB 5f6572f1e3f9
coco/consensus_nce_mse_s0_t8000.ptWeights34.7 MB 090e5bba6ffd
coco/consensus_nce_mse_s1_t8000.ptWeights34.7 MB 9ac63b8f2d04
coco/feature_mse_s0_t8000.ptWeights34.7 MB 774080c231b6
coco/feature_mse_s1_t8000.ptWeights34.7 MB 8c2ee5d7e31e
coco/infonce_s0_t8000.ptWeights34.7 MB ec258fdc7e02
coco/infonce_s1_t8000.ptWeights34.7 MB 667628dafa81
coco/nce_mse_cv_s0_t8000.ptWeights34.7 MB 4668acc5db17
coco/nce_mse_cv_s1_t8000.ptWeights34.7 MB 5acf925fb4f2
coco/siglip_pairwise_s0_t8000.ptWeights35.0 MB 6d972139a255
coco/siglip_pairwise_s1_t8000.ptWeights35.0 MB 5f68e994daf9
coco/x3_autograd_s0_t8000.ptWeights34.7 MB 9191b7fa5aa8
coco/x3_autograd_s1_t8000.ptWeights34.7 MB f57ccbf5ccc7
coco/x3_bce_s0_t8000.ptWeights34.7 MB b16ba4837758
coco/x3_bce_s1_t8000.ptWeights34.7 MB 53b3917b1930
coco/x3_full_s0_t8000.ptWeights34.7 MB c13544b790d4
coco/x3_full_s1_t8000.ptWeights34.7 MB 79f083639621
model.safetensorsWeights35.7 MB 82c7102e1c7e
splat_screen/feature_mse_s4_splat_all_s0_t8000.ptWeights34.7 MB 1ff106c2976c
splat_screen/feature_mse_s4_splat_all_s1_t8000.ptWeights34.7 MB ad44cd2ccbd9
splat_screen/infonce_rep1_s0_t8000.ptWeights34.7 MB a1ab260ca3fe
splat_screen/infonce_s2_s0_t2000.ptWeights34.7 MB 5d3370bd3240
splat_screen/infonce_s2_splat_all_s0_t2000.ptWeights34.7 MB 6107833f1a6f
splat_screen/infonce_s3_splat_all_frozen_s0_t2000.ptWeights34.7 MB 41ece178371a
splat_screen/infonce_s3_splat_b12_s0_t2000.ptWeights34.7 MB 025c7a567039
splat_screen/infonce_s3_splat_b1_3_s0_t2000.ptWeights34.7 MB 92b10eb68619
splat_screen/infonce_s4_splat_all_frozen_s0_t8000.ptWeights34.7 MB 492687e36399
splat_screen/infonce_s4_splat_all_frozen_s1_t8000.ptWeights34.7 MB 821d6fc860de
splat_screen/infonce_s4_splat_all_s0_t8000.ptWeights34.7 MB 0fe2aaf5abec
splat_screen/infonce_s4_splat_all_s1_t8000.ptWeights34.7 MB d1572d5b2983
splat_screen/infonce_s4_splat_b1_6_s0_t8000.ptWeights34.7 MB 9ac20531ff57
splat_screen/infonce_s4_splat_b1_6_s1_t8000.ptWeights34.7 MB 56c4b577a785
splat_screen/infonce_s4_splat_b7_12_s0_t8000.ptWeights34.7 MB f4f781891e51
splat_screen/infonce_s4_splat_b7_12_s1_t8000.ptWeights34.7 MB 3b51e7d0e26b
splat_screen/infonce_s5_ls_fixed1_s0_t2000.ptWeights34.7 MB 2d1aec61a016
splat_screen/infonce_s5_ls_fixed1_s1_t2000.ptWeights34.7 MB 2c9779493af4
splat_screen/infonce_s5_ls_learn_s0_t2000.ptWeights34.7 MB b666bfdba09b
splat_screen/infonce_s5_ls_learn_s0_t8000.ptWeights34.7 MB 9a8e0ff7ad88
splat_screen/infonce_s5_ls_learn_s1_t2000.ptWeights34.7 MB 2e774544b74b
splat_screen/infonce_s5_ls_learn_s1_t8000.ptWeights34.7 MB ac62e61740f1
splat_screen/infonce_s5_ls_zero_s0_t2000.ptWeights34.7 MB 31d04633c166
splat_screen/infonce_s5_ls_zero_s1_t2000.ptWeights34.7 MB 79346e899cd5
student_infonce_s0.ptWeights34.7 MB ec258fdc7e02
config.jsonConfiguration487 B —
configuration_clip_mini.pyConfiguration1.2 KB —
ledgers/coco_crossfit_2026-10-08.jsonConfiguration1.7 KB —
ledgers/coco_heldout_rescore_2026-10-08.jsonConfiguration2.0 KB —
ledgers/extra_ceiling.jsonConfiguration106 B —
ledgers/frame_check.jsonConfiguration343 B —
ledgers/frame_check_pod_s0.jsonConfiguration767 B —
ledgers/frame_check_pod_s1.jsonConfiguration763 B —
ledgers/frame_check_s1.jsonConfiguration347 B —
ledgers/rotated_eval_consensus_gpa.jsonConfiguration292 B —
ledgers/rotated_eval_consensus_nce_mse.jsonConfiguration290 B —
loader.pyConfiguration4.2 KB —
modeling_clip_mini.pyConfiguration4.0 KB —
preprocessor_config.jsonConfiguration457 B —
splat_screen/probe_init_sdpa_s0.jsonConfiguration1.3 KB —
splat_screen/probe_init_splat_s0.jsonConfiguration2.6 KB —
splat_screen/probe_ls_capacity_infonce_s4_splat_all_s0_t8000.jsonConfiguration150.9 KB —
splat_screen/probe_ls_capacity_infonce_s4_splat_b7_12_s0_t8000.jsonConfiguration75.6 KB —
train/bank_utils.pyConfiguration1.9 KB —
train/cc12m_data.pyConfiguration6.1 KB —
train/cc12m_gpa.pyConfiguration3.6 KB —
train/dist_align_gate.pyConfiguration16.3 KB —
train/dist_bed.pyConfiguration41.2 KB —
train/losses.pyConfiguration6.2 KB —
train/vitals.pyConfiguration1.6 KB —
README.mdDocumentation13.9 KB —
article_cc12m_distillation.mdDocumentation28.8 KB —
train/TRAINING.mdDocumentation6.6 KB —
ledgers/affinity_kl_s0.jsonlOther1.1 KB —
ledgers/affinity_kl_s1.jsonlOther1.1 KB —
ledgers/blueprint_s0.jsonlOther708 B —
ledgers/blueprint_s1.jsonlOther702 B —
ledgers/cbert_full_s0.jsonlOther1.1 KB —
ledgers/cbert_full_s1.jsonlOther1.1 KB —
ledgers/consensus_gpa_s0.jsonlOther1.1 KB —
ledgers/consensus_gpa_s1.jsonlOther1.1 KB —
ledgers/consensus_nce_mse_s0.jsonlOther1.1 KB —
ledgers/consensus_nce_mse_s1.jsonlOther1.1 KB —
ledgers/feature_mse_s0.jsonlOther1.4 KB —
ledgers/feature_mse_s1.jsonlOther1.4 KB —
ledgers/infonce_s0.jsonlOther1.8 KB —
ledgers/infonce_s1.jsonlOther1.4 KB —
ledgers/nce_mse_cv_s0.jsonlOther1.1 KB —
ledgers/nce_mse_cv_s1.jsonlOther1.1 KB —
ledgers/siglip_pairwise_s0.jsonlOther1.1 KB —
ledgers/siglip_pairwise_s1.jsonlOther1.1 KB —
ledgers/teacher_ceiling_s0.jsonlOther317 B —
ledgers/x3_autograd_s0.jsonlOther1.1 KB —
ledgers/x3_autograd_s1.jsonlOther1.1 KB —
ledgers/x3_bce_s0.jsonlOther404 B —
ledgers/x3_bce_s1.jsonlOther403 B —
ledgers/x3_full_s0.jsonlOther407 B —
ledgers/x3_full_s1.jsonlOther407 B —
ledgers/zs_floor_s0.jsonlOther317 B —
splat_screen/feature_mse_s4_splat_all_s0.jsonlOther4.7 KB —
splat_screen/feature_mse_s4_splat_all_s1.jsonlOther4.7 KB —
splat_screen/infonce_evalchk_s0.jsonlOther361 B —
splat_screen/infonce_rep1_s0.jsonlOther2.4 KB —
splat_screen/infonce_s2_s0.jsonlOther2.4 KB —
splat_screen/infonce_s2_splat_all_s0.jsonlOther4.7 KB —
splat_screen/infonce_s3_splat_all_frozen_s0.jsonlOther4.7 KB —
splat_screen/infonce_s3_splat_b12_s0.jsonlOther2.6 KB —
splat_screen/infonce_s3_splat_b1_3_s0.jsonlOther3.0 KB —
splat_screen/infonce_s4_splat_all_frozen_s0.jsonlOther4.7 KB —
splat_screen/infonce_s4_splat_all_frozen_s1.jsonlOther4.7 KB —
splat_screen/infonce_s4_splat_all_s0.jsonlOther4.7 KB —
splat_screen/infonce_s4_splat_all_s1.jsonlOther4.7 KB —
splat_screen/infonce_s4_splat_b1_6_s0.jsonlOther3.5 KB —
splat_screen/infonce_s4_splat_b1_6_s1.jsonlOther3.5 KB —
splat_screen/infonce_s4_splat_b7_12_s0.jsonlOther3.6 KB —
splat_screen/infonce_s4_splat_b7_12_s1.jsonlOther3.6 KB —
splat_screen/infonce_s5_ls_fixed1_s0.jsonlOther8.0 KB —
splat_screen/infonce_s5_ls_fixed1_s1.jsonlOther4.0 KB —
splat_screen/infonce_s5_ls_learn_s0.jsonlOther12.2 KB —
splat_screen/infonce_s5_ls_learn_s1.jsonlOther8.2 KB —
splat_screen/infonce_s5_ls_zero_s0.jsonlOther4.0 KB —
splat_screen/infonce_s5_ls_zero_s1.jsonlOther4.0 KB —
.gitattributesRepository1.5 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
2.9 GB
Download from AbstractPhila

Released by AbstractPhila through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published2.9 GB
16-bit0.0 GB
8-bit0.0 GB
4-bit0.0 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About clip-vitb-mini-distilled

How much GPU memory does clip-vitb-mini-distilled need?

About 0 GB at 16-bit and 0 GB at 4-bit: the weights (9M parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run clip-vitb-mini-distilled on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use clip-vitb-mini-distilled commercially?

Yes. clip-vitb-mini-distilled is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Image feature extraction

I-JEPA-Agro

Junior R F Junior

Projeto completo de triagem visual de anomalias em folhas de soja usando como extrator de representações e HeraclitusDB como banco de eventos e vetores. 1. lê um manifesto com fazenda, talhão, safra e cultivar; 2. executa o I-JEPA ViT-G/16 localmente; 3. aplica média aos patch tokens e normalização L2; 4. persiste cada imagem e seu embedding euclidiano no HeraclitusDB; 5. cria um artefato local para avaliação offline; 6. treina uma sonda linear; 7. busca imagens visualmente semelhantes; 8. oferece busca por uma API FastAPI. - Python 3.11+; - HeraclitusDB ativo em 127.0.0.1:7474; - vários GB livres para o checkpoint; - RAM suficiente para um modelo com aproximadamente 1B parâmetros; - GPU é…

Open weights mit transformers
View model

Model · Image feature extraction

IM-Animation-Motion-Encoder

Fynn

IM-Animation: An Implicit Motion Representation for Identity-decoupled Character Animation This release contains the final locally retained TiTok-based motion encoder implementation and an exported checkpoint. It provides frame-level motion tokens; it does not include the full animation generator or retargeting network. Code and project videos are available in the linked GitHub repository. The encoder was verified with PyTorch 2.7.1 on CPU using the exported BF16 weights. The weight file is 607,017,208 bytes (approximately 579 MiB). Its SHA256 and provenance are in checkpointinfo.json. frames must be RGB, normalized to [0, 1], with shape [N, 3, 256, 256]. The training preprocessing pads…

Open weights apache-2.0
View model

Model · Image feature extraction

RK182X-VIT-siglip2-so400m-patch14-384

RKNNAI

本仓库提供由 google/siglip2-so400m-patch14-384 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VIT-siglip2-so400m-patch14-384 - 模型显示名称：RK182X-VIT-siglip2-so400m-patch14-384 - 源模型：google/siglip2-so400m-patch14-384 - 模型类型：VIT ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 源模型许可证：Apache License 2.0，正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。

Open weights apache-2.0
View model

Model · Image feature extraction

RK182X-VIT-siglip-so400m-patch14-384

RKNNAI

本仓库提供由 google/siglip-so400m-patch14-384 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VIT-siglip-so400m-patch14-384 - 模型显示名称：RK182X-VIT-siglip-so400m-patch14-384 - 源模型：google/siglip-so400m-patch14-384 - 模型类型：VIT ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 源模型许可证：Apache License 2.0，正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。

Open weights apache-2.0
View model

Model · Image feature extraction

RK182X-VIT-dinov3-vits16-pretrain-lvd1689m

RKNNAI

本仓库提供由 facebook/dinov3-vits16-pretrain-lvd1689m 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VIT-dinov3-vits16-pretrain-lvd1689m - 模型显示名称：RK182X-VIT-dinov3-vits16-pretrain-lvd1689m - 源模型：facebook/dinov3-vits16-pretrain-lvd1689m - 模型类型：VIT ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 源模型许可证：DINOv3 License，正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。

Open weights other
View model

Model · Image feature extraction

RK182X-VIT-dinov3-vitl16-pretrain-lvd1689m

RKNNAI

本仓库提供由 facebook/dinov3-vitl16-pretrain-lvd1689m 转换的 RKNN 模型。 - Model ID：RKNNAI/RK182X-VIT-dinov3-vitl16-pretrain-lvd1689m - 模型显示名称：RK182X-VIT-dinov3-vitl16-pretrain-lvd1689m - 源模型：facebook/dinov3-vitl16-pretrain-lvd1689m - 模型类型：VIT ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 源模型许可证：DINOv3 License，正文见 LICENSE，归属与转换修改说明见 NOTICE。 - 同时遵守 RKNN Toolkit、RKNN Runtime 相关许可；本包不包含运行库。

Open weights other
View model