NEWVectors or files. Pick a path.Start →

    AI Model Hub

    Browse AI models for multimodal decomposition and recomposition pipelines: plug any model into your extractors.

    19,428 models available

    Run models on your own files, freeSchedule DemoRead Documentation

    Showing 217-240 of 19,428 models

    Image Text To Text

    Qwen/Qwen3.5-35B-A3B

    1.5M
    1,546
    transformers
    View Details
    Automatic Speech Recognition

    Khalsuu/filipino-wav2vec2-l-xls-r-300m-official

    1.5M
    14
    transformers
    View Details
    Depth Estimation

    depth-anything/Depth-Anything-V2-Small-hf

    1.5M
    69
    transformers
    View Details
    Text To Image

    stable-diffusion-v1-5/stable-diffusion-v1-5

    1.5M
    1,086
    diffusers
    View Details
    Text Generation

    Qwen/Qwen2.5-Coder-7B-Instruct

    1.5M
    844
    transformers
    View Details
    Zero Shot Image Classification

    openai/clip-vit-base-patch16

    1.5M
    182
    transformers
    View Details
    Audio To Audio

    nvidia/bigvgan_v2_22khz_80band_256x

    1.4M
    39
    PyTorch
    View Details
    Automatic Speech Recognition

    kingabzpro/wav2vec2-large-xls-r-300m-Urdu

    1.4M
    22
    transformers
    View Details
    Text Generation

    nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

    1.4M
    835
    transformers
    View Details
    Automatic Speech Recognition

    jonatasgrosman/wav2vec2-large-xlsr-53-finnish

    1.4M
    4
    transformers
    View Details
    Text Generation

    deepseek-ai/DeepSeek-V3

    1.4M
    4,451
    transformers
    View Details
    Image Text To Text

    Qwen/Qwen3-VL-8B-Instruct-FP8

    1.4M
    94
    transformers
    View Details
    Text Generation

    ornith-ai/Ornith-1.0-35B

    1.4M
    518
    transformers
    View Details
    Fill Mask

    google-bert/bert-base-multilingual-cased

    1.4M
    616
    transformers
    View Details
    Token Classification

    StanfordAIMI/stanford-deidentifier-base

    1.4M
    93
    transformers
    View Details
    Image Text To Text

    Qwen/Qwen2-VL-7B-Instruct-AWQ

    1.4M
    56
    transformers
    View Details
    Sentence Similarity

    nomic-ai/nomic-embed-text-v1

    1.4M
    596
    sentence-transformers
    View Details
    Sentence Similarity

    sentence-transformers/paraphrase-MiniLM-L6-v2

    1.4M
    161
    sentence-transformers
    View Details
    Image Text To Text

    unsloth/Qwen3.5-9B-GGUF

    1.4M
    1,033
    transformers
    View Details
    Feature Extraction

    cambridgeltl/SapBERT-from-PubMedBERT-fulltext

    1.4M
    88
    transformers
    View Details
    Automatic Speech Recognition

    mesolitica/wav2vec2-xls-r-300m-mixed

    1.4M
    7
    transformers
    View Details
    Automatic Speech Recognition

    nguyenvulebinh/wav2vec2-base-vi-vlsp2020

    1.4M
    13
    transformers
    View Details
    Image Text To Text

    datalab-to/surya-ocr-2

    1.4M
    139
    transformers
    View Details
    Audio Classification

    MIT/ast-finetuned-audioset-10-10-0.4593

    1.4M
    384
    transformers
    View Details
    10 / 810

    Choosing a model for multimodal retrieval

    Which of these models does Mixpeek actually run?

    Nine, and they are not the same thing as the catalog. The managed extractors run intfloat/multilingual-e5-large-instruct for text (1024-d), google/siglip-base-patch16-224 for images (768-d), google/vertex-multimodal (1408-d) and google/gemini-embedding-2 (3072-d) for unified multimodal, insightface ArcFace for faces (512-d), CLAP for audio fingerprints (512-d), facebook/dinov2-base for visual similarity and jinaai/jina-embeddings-v2-base-code for code (both 768-d, inside the web scraper). The rerank retriever stage runs BAAI/bge-reranker-v2-m3. Everything else in this catalog is documented here, not hosted here.

    Can I run any model from this catalog on Mixpeek?

    Not by naming it. No extractor takes a Hugging Face model id as a parameter, so there is no field to put one in. Three paths do work. Use a managed extractor and get the model it runs. Run the model on your own hardware and upsert the vectors through POST /v1/namespaces/{namespace_id}/documents/upsert, which stores them beside everything else. Or on Enterprise, upload the weights through POST /v1/namespaces/{namespace_id}/models, which accepts the huggingface format, and load them from a custom plugin.

    Should I use a text embedding model or a multimodal one?

    Ask whether a text query has to reach a non-text asset directly. If your video is searchable through its transcript and your images through their captions, a text model over that generated text is cheaper and usually more accurate, because retrieval quality on words is a solved problem and cross-modal alignment is not. If the query is 'find the shot that looks like this' or the visual content carries meaning no caption records, you need a shared space and a multimodal model. Most production systems run both indexes rather than choosing.

    What does the embedding dimension cost to store?

    A float32 vector is 4 bytes per dimension, so a million items costs 4 GB at 1024 dimensions, 3 GB at 768, and 12 GB at 3072. That is before any quantization and before payload. Models trained with Matryoshka representation learning, such as Gemini Embedding 2, let you truncate to a shorter prefix without re-encoding, so the width becomes an index-time decision rather than a model-selection one. Dimensions are fixed at namespace creation in Mixpeek, so changing a width later means re-indexing.

    Why does the download count on a model page differ from HuggingFace?

    It is the monthly figure from the HuggingFace API at the time of the last sync, not a live read, so it lags. The sync date is on each model page. Download count is a popularity signal and a poor quality signal: the most-downloaded model in a category is frequently an older checkpoint that a tutorial pinned years ago.

    Previous

    Connectors

    Next

    Retrievers

    What's new in models

    Full changelog

    All models

    Every model page in one place, 442 in total.