All 82 models, by task
Mozo serves 82 variants across 18 families. A model is named family/variant, the same in Python
(mozo.get_model("rfdetr/medium")) and over HTTP (/predict/rfdetr/medium). Each family's page
has examples, every parameter, and how the model behaves.
Sizes are the torch weights a first load downloads. The runtimes are the ones mozo can execute for
that family; runtime="auto" picks among them for your device.
Object detection
A box, a class and a confidence for each object.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| RF-DETR | nano small medium large |
Apache-2.0 | torch, onnx, coreml | 136–405 MB |
| YOLOv8 | nano small medium large xlarge |
AGPL-3.0 | torch, onnx, coreml | 7–137 MB |
| YOLO11 | nano small medium large xlarge |
AGPL-3.0 | torch, onnx | 6–115 MB |
| YOLO12 | nano small medium large xlarge |
AGPL-3.0 | torch, onnx, coreml | 6–119 MB |
| YOLO26 | nano small medium large xlarge |
AGPL-3.0 | torch, onnx | 6–119 MB |
Instance segmentation
A box and a mask for each object, in one pass. Same call as detection: the variant decides whether masks come back.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| RF-DETR | seg-nano seg-small seg-medium seg-large |
Apache-2.0 | torch, onnx, coreml | 135–145 MB |
| YOLO11 | seg-nano seg-small seg-medium seg-large seg-xlarge |
AGPL-3.0 | torch, onnx | 6–125 MB |
| YOLO26 | seg-nano seg-small seg-medium seg-large seg-xlarge |
AGPL-3.0 | torch, onnx | 7–142 MB |
Keypoints
COCO's 17 joints for each person, in pixels of your image. RF-DETR finds people and their joints in one pass; ViTPose adds joints to boxes you already have.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| RF-DETR | keypoint-preview |
Apache-2.0 | torch | 164 MB |
| ViTPose | small base large huge |
Apache-2.0 | torch, onnx (not huge) |
133 MB–3.6 GB |
Detection from words
Boxes for anything you can name, from a phrase instead of a fixed class list.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| OWLv2 | base base-ensemble large large-ensemble |
Apache-2.0 | torch | 620 MB–1.8 GB |
| Grounding DINO | tiny base |
Apache-2.0 | torch | 694–938 MB |
| SAM 3 | sam3 (masks as well as boxes) |
SAM License | torch, coreml | 3.5 GB |
Segmentation from a click or a box
Point at something and get a mask of the thing you pointed at.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| SAM 2 | tiny small base_plus large |
Apache-2.0 | torch | 156–898 MB |
| EdgeTAM | edgetam |
Apache-2.0 | torch | 56 MB |
| SAM 3 | sam3 |
SAM License | torch, coreml | 3.5 GB |
Classification from words, and image embeddings
Name your classes in words and each one is scored against the image, or take the vectors to search images by text in a database of your own.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| CLIP | base base-16 large large-336 |
MIT | torch | 599 MB–1.7 GB |
| SigLIP 2 | base-224 base-256 so400m-384 so400m16-256 giant-384 |
Apache-2.0 | torch | 1.5–7.5 GB |
Text embeddings
English text in, one vector out, for searching documents by meaning. Text only: these vectors do not share a space with CLIP's or SigLIP 2's.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| Granite Embedding R2 | small |
Apache-2.0 | torch | 191 MB |
Text recognition
Every line of text in the image, read. A variant is a script, not a language: latin covers 41
languages.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| EasyOCR | english latin chinese-simplified japanese korean |
Apache-2.0 | torch | 98–105 MB |
Depth
A depth map at the image's resolution, in metres or relative.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| Depth Anything V2 | small (relative) |
Apache-2.0 | torch | 99 MB |
base large (relative) |
CC-BY-NC-4.0 | torch | 390 MB–1.3 GB | |
indoor- and outdoor- × small base large (metres) |
Apache-2.0 | torch | 99 MB–1.3 GB |
Background removal
An opacity matte for the foreground, soft enough to keep hair.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| BEN2 | base |
MIT | torch | 381 MB |
Object removal
An image and a mask in, the masked thing painted out.
| Family | Variants | Licence | Runtimes | Size |
|---|---|---|---|---|
| Moebius | general places2 |
Apache-2.0 | torch | 1.1 GB |
Ask mozo instead of this page
The same facts are available in code, without loading a model or touching the network:
import mozo
from mozo import weights
mozo.MODEL_REGISTRY["rfdetr"]["variants"] # ['nano', 'small', ..., 'keypoint-preview']
weights.license_of("yolov8", "nano") # 'AGPL-3.0'Over HTTP, GET /models lists every model the server offers, with its task and licence.