Mozo

63 computer vision models. One pip install. No dependency hell.

Normally each of these models arrives with its own package, and each package brings its own dependencies — torch, numpy and OpenCV, every one pinned to something slightly different. Put a few in one environment and something breaks. The usual escape is a container per model, and paying for that forever.

Mozo ships none of them. Every model's inference path is vendored into mozo itself and verified bit-identical to the original implementation — so one environment runs all 63, and gives you the original's exact numbers rather than something close.

pip install mozo
mozo start

No Docker, no Kubernetes, no per-model build step, no conversion step, no git clone of anything.

Serve from one install, in one process, behind one API.

Model Catalog

Object detection

Returns detections.

Text-prompted

Returns detections.

Zero-shot classification and embeddings

Name your classes in words and each is scored against the image. The same model will also hand back the vectors it works from, which is what makes a corpus embedded once searchable by words afterwards — through a vector database of your own.

Scores are cosine similarities, not probabilities: not softmaxed, they do not sum to one, and they may be negative. Nothing is filtered out — every phrase comes back scored, because a classifier that drops a class has not classified.

The only family with a second route. POST /encode/clip/base returns the embeddings instead of an answer, for images or for phrases; the towers load independently, so a job that only encodes images never allocates the text half.

Promptable segmentation

Returns detections, one row per mask.

Text recognition

A variant is a script, not a language: latin alone covers 41 languages.

Returns detections, one row per line.

Depth

Returns an HxW float32 map.

Relative output is inverse depth on a per-image scale: larger is nearer, and no value is a distance. model.unit says which, and is None rather than a guess.

How each family behaves and why — OWLv2 suppressing nothing, the encoder cache, EasyOCR's line ordering, prompt semantics — is in docs/models.md.

Install

pip install mozo

That is the whole install — twelve ordinary dependencies, not one of them a model package.

# Optional runtimes, for the families that publish those artifacts
pip install 'mozo[onnx]'     # onnxruntime
pip install 'mozo[coreml]'   # coremltools, macOS only

No family needs an extra. The extras add ways to execute a model, not models: without them runtime="auto" simply does not select those artifacts.

Using it

Two interfaces over the same models, plus a browser page for testing them.

1. HTTP server

mozo start

curl -X POST "http://localhost:8000/predict/rfdetr/medium" -F "file=@street.jpg"

[
  {"bbox": [0.0, 113.24, 763.19, 1269.33], "class_name": "person", "confidence": 0.917},
  {"bbox": [709.0, 773.0, 1366.0, 1143.0], "class_name": "laptop", "confidence": 0.854}
]

Trimmed for reading. The real response carries every PixelFlow field on every detection — masks, segments, text, class_id, metadata and the rest — including the ones that are null, because masks: null beside a filled segments is a difference worth being able to see.

The catalogue is answerable without loading anything:

curl http://localhost:8000/models          # all 63, no torch import, no weights
curl http://localhost:8000/models/loaded   # what is resident right now

Full parameter reference below, and the server documents itself at http://localhost:8000/docs.

2. Python API

import mozo

model = mozo.get_model("rfdetr/medium")

for found in model.predict("street.jpg"):
    print(found.class_name, found.confidence, found.bbox)
#   person 0.92 [0.0, 113.24, 763.19, 1269.33]
#   laptop 0.85 [709.0, 773.0, 1366.0, 1143.0]

predict takes a path, encoded bytes, or an RGB array, and returns a unified PixelFlow Detections — iterable, indexable, and the same shape from every family here.

# Ask for something in words
model = mozo.get_model("owlv2/base-ensemble")
found = model.predict("kitchen.jpg", ["kettle", "a mug", "the window"])
found[0].class_name          # 'kettle' — the phrase you searched for

# Read the text on a sign
model = mozo.get_model("easyocr/english")
found = model.predict("sign.jpg")
found[0].text                # 'EXIT 42' — what it says
found[0].class_name          # None — OCR reads content, it does not pick a class

# Measure depth
model = mozo.get_model("depth_anything_v2/indoor-small")
depth = model.predict("room.jpg")    # HxW float32, at the input's resolution
model.unit                           # 'metres' — or None, and then it is not a distance

Try it in the browser

mozo start        # then open http://localhost:8000/test-ui

Pick any of the 63, run it on your own image, and see the response two ways at once: drawn on the image, and as the raw PixelFlow record. Hovering a box lights its row and its JSON, so when something lands somewhere surprising its numbers are one click away.

Verification

Vendoring a model normally means maintaining something that quietly drifts from the original. The whole claim above rests on that not happening, so it is checked rather than asserted — with no tolerance. Exact equality, because a tolerance hides precisely the drift a check exists to catch.

The gates in tools/verify/ compare every intermediate stage against the original implementation, not just the final answer: 1,275 comparisons for EasyOCR, 226 for OWLv2, 138 for Grounding DINO, every one identical. Twelve of the fourteen families ship one; all fourteen have their parity measured and recorded in mozo/vendors/<family>_deploy/PROVENANCE.md, with the upstream commit it was built from.

Because the extraction is the implementation rather than a wrapper around one, none of this costs anything at run time: EasyOCR runs within 1% of the published package on the same weights.

HTTP API

POST /predict/{family}/{variant}
Content-Type: multipart/form-data

text is deliberately not comma-separated the way labels is: a prompt is free text, so ?text=a person, holding a mug stays one concept rather than becoming two wrong ones. label is required with point and has no default — guessing between include and exclude returns a confident mask of the wrong thing.

# Boxes for anything you can name, under a permissive licence
curl -X POST "http://localhost:8000/predict/owlv2/base-ensemble?text=kettle&text=a%20mug" \
  -F "file=@kitchen.jpg"

# Click one thing: three candidate masks, best first
curl -X POST "http://localhost:8000/predict/edgetam/edgetam?point=820,640&label=1" \
  -F "file=@street.jpg"

# A box, one mask, named by you
curl -X POST "http://localhost:8000/predict/sam2/tiny?box=40,60,300,480&multimask=false&name=cat" \
  -F "file=@street.jpg"

Depth answers with an image rather than JSON:

curl -X POST "http://localhost:8000/predict/depth_anything_v2/indoor-small" \
  -F "file=@room.jpg" -D headers.txt --output depth.png

# depth.png is a 16-bit PNG; headers.txt carries what it means:
#   X-Depth-Unit: metres | none      X-Depth-Min / X-Depth-Max: the endpoints
#   depth = min + png / 65535 * (max - min)

16-bit rather than 8, because six of the nine variants predict metres and quantising those to 256 levels would discard the measurement.

CLIP is the one family with a second route. /predict classifies; /encode returns the vectors instead of an answer, for images or for phrases but not both in one call:

# Score an image against phrases you make up
curl -X POST "http://localhost:8000/predict/clip/base?text=a%20forklift&text=a%20person" \
  -F "file=@aisle.jpg"

# The vectors themselves, to put in a vector database of your own
curl -X POST "http://localhost:8000/encode/clip/base" -F "file=@aisle.jpg"
curl -X POST "http://localhost:8000/encode/clip/base?text=a%20forklift&text=a%20person"

The response carries model and revision alongside the embeddings, and those are not decoration: a vector is only comparable against others from the same weights, so a stored index is tied to them.

The other endpoints: GET / for health and residency, GET /models for the catalogue, GET /models/loaded for what is in memory, GET /test-ui for the browser page, and GET /docs for the generated OpenAPI reference.

Python API

mozo.get_model uses one process-wide cache. When you want a separate lifetime — a batch job that should release its models at the end — build your own manager and drop it:

from mozo import ModelManager

models = ModelManager()
model = models.get_model("rfdetr", "medium", device="cpu")

scratch = ModelManager()      # a separate lifetime; drop it and its models go with it

device takes "cuda", "mps", "cpu", or None to take the best available.

Your own checkpoints

Fine-tuned weights load through the same API, on architectures mozo supports:

model = models.get_model(
    "rfdetr", "my-training",
    checkpoint_path="runs/best.pth",
    model_size="small", project_type="detection",
    labels=["hardhat", "vest"],
)

Everything that changes what gets built is part of the cache identity — a checkpoint, a pinned revision, a runtime — so two people's my-training are two models rather than one.

PixelFlow

Every detection family returns the same object, so filtering and annotation are written once:

import pixelflow as pf

found = model.predict(image)
filtered = found.filter_by_confidence(0.8).filter_by_class_id([0, 2])

annotated = pf.annotate.box(image, filtered)
annotated = pf.annotate.label(annotated, filtered)

json_output = filtered.to_json()

More: PixelFlow

How it works

Nothing loads until it is asked for. The server starts instantly whatever is published, and the first request for a family downloads and loads it. That first request is slow — minutes, for a multi-gigabyte family on a cold cache — and every one after it is not.

Nothing is evicted. A model stays for the life of the process. This is deliberate: an earlier version bounded the cache by model count, and a count cannot tell 0.10 GB from 1.34 GB. The same number could not be right for a 6 GB laptop and an 80 GB accelerator, and measured, it broke the obvious deployment — detection, segmentation and depth from one instance, 0.60 GB between them — into a 762 ms eviction on every request. Memory is yours to manage, and the lever is which models you ask for.

A model is built once however many requests arrive for it at the same moment, and a request for a model already in memory never waits behind an unrelated load.

Weights resolve from a manifest that ships inside the package, so working out which bytes a model refers to needs no network and no configuration. Every artifact is verified by sha256 after download, and its licence and a NOTICE naming the exact upstream release are published beside it.

Configuration

Deploying

mozo start                      # 0.0.0.0:8000
mozo start --port 8080
mozo start --workers 4          # read the note below first
mozo start --reload             # development; forces one worker

It is a real server — FastAPI on uvicorn, thread-safe, with weights verified by hash — and three things about it are yours to arrange:

• --workers N multiplies memory by N. Workers are separate processes and share nothing, so each loads its own copy of every model it serves. Four workers serving SAM 3 is four times 3.4 GB, not one. Prefer one worker unless you have measured otherwise.
• There is no authentication and no rate limiting, and the default bind is 0.0.0.0. Put it behind something before it faces a network you do not control.
• Nothing is evicted, so a process asked for every family will eventually hold every family. Decide what an instance serves rather than letting callers decide for it — MOZO_ENABLE is how, and it is also how the licence question below is answered.

What mozo does not do

• No training and no fine-tuning. Bring a checkpoint; mozo runs it.
• No video, no tracking, no streams. One image per request.
• No batching. One image per forward, which is what keeps results bit-identical.
• No model conversion. ONNX and CoreML artifacts are published where a family exports cleanly, and where it does not, mozo says so rather than shipping a graph that disagrees.
• It is not a model hub. The catalogue is a curated 63, chosen because each one could be extracted and verified. Growth is deliberate and slow.

Extending

1. Write an adapter in mozo/adapters/your_model.py
2. Register it in mozo/registry.py
3. It is available over HTTP, in Python, and in the test UI

HTTP request → FastAPI server → ModelManager → Adapter → Vendor
                                     ↓
                               thread-safe cache

The rule that keeps this honest: a vendor imports no other vendor and never imports mozo. Duplication between vendors is deliberate, so a family can be re-extracted from a newer upstream release without touching anything else. tests/test_vendor_agreement.py enforces it, and each vendor's PROVENANCE.md records the exact upstream commit it came from and what was changed.

Development

pip install -e .
mozo start --reload
pytest

Links

• Repository
• Issues
• PixelFlow — the result format

License

Mozo's own code is Apache-2.0, and so is every vendored extraction under mozo/vendors/.

The weights are separate works travelling with it. Of the 63 published variants, 36 are Apache-2.0, 20 are AGPL-3.0 (every YOLO variant), 4 are MIT (CLIP), 2 are CC-BY-NC-4.0 (Depth Anything base and large), and 1 carries Meta's SAM License (SAM 3). The full licence and a NOTICE naming the exact upstream release are published beside every checkpoint.

YOLO weights are AGPL-3.0, or covered by a commercial licence from Ultralytics. And serving predictions from them over a network places AGPL-3.0 section 13 obligations on you.

Complying with either is the operator's responsibility. MOZO_ENABLE is how an instance declines to take it on — a server that never offers a model never fetches or serves its weights:

MOZO_ENABLE=clip,easyocr,edgetam,grounding_dino,owlv2,rfdetr,sam2,siglip2,\
depth_anything_v2/small,depth_anything_v2/indoor-small,depth_anything_v2/indoor-base,\
depth_anything_v2/indoor-large,depth_anything_v2/outdoor-small,\
depth_anything_v2/outdoor-base,depth_anything_v2/outdoor-large

That is the 40 Apache-2.0 and MIT variants, and it stays 40 through an upgrade that adds families — which a list of what to exclude would not. Depth Anything is named variant by variant because it is the one family whose licence is not uniform: seven of its nine are Apache-2.0 and two are CC-BY-NC-4.0.

Nothing here relicenses anything. It only decides what one deployment hands out.