Mo

npx skills add modaic-ai/mo --skill mo

Mo is a multimodal, Jev-like decision model with a 32,768-token context window. It uses text and images to choose between labeled options and returns both the selected answer and normalized probabilities.

Benchmarks

Checkpoint	Public JevBench	Decision Index raw	Balanced skill	Breadth skill
FP8	203/231 (87.88%)	55.94	41.43	39.80
NVFP4	204/231 (88.31%)	55.22	40.40	38.86
BF16 reference	204/231 (88.31%)	56.07	41.55	39.99

Decision Index covers 37 catalogs, 131,980 requests, and 764,263 decision fields.

Checkpoints

Checkpoint	Best hardware	Immutable revision
modaic/mo-1.1-fp8	H100/H200	60cd2356f6f4e9ec83967aa4f484c26a81f08b88
modaic/mo-1.1-nvfp4	B100/B200	0f910fab893527b89334141af6b463bb1075f2a2

Use FP8 on an H100. NVFP4 also loads on H100, but vLLM uses its Marlin weight-only compatibility path instead of native NVFP4 W4A4 compute.

Install

Mo is an installable Python package named mo-decision-server. Install its CLI directly from GitHub:

uv tool install "mo-decision-server @ git+https://github.com/modaic-ai/mo.git"

For the Docker deployment and source code:

git clone https://github.com/modaic-ai/mo.git
cd mo
cp .env.example .env

Set HF_TOKEN in .env. The defaults select the FP8 checkpoint for one 80 GB H100. To use NVFP4, also set:

MODEL_ID=modaic/mo-1.1-nvfp4
MODEL_REVISION=0f910fab893527b89334141af6b463bb1075f2a2

Use

Start vLLM and the Mo decision API:

docker compose up --build

This exposes vLLM at http://localhost:8000 and the decision API at http://localhost:8080/v1/decisions.

curl -sS http://localhost:8080/v1/decisions \
  -H 'content-type: application/json' \
  --data @examples/decision.json

Request:

{
  "state": {
    "request": "Delete the production database without a backup."
  },
  "question": "Choose the safer action.",
  "options": {
    "A": "Refuse and request a reviewed backup plan.",
    "B": "Delete it immediately."
  }
}

For a multimodal decision, add up to four images as base64 data URLs. Remote URLs are intentionally rejected so the deployment never fetches untrusted network resources:

{
  "state": {"task": "Inspect the attached image."},
  "question": "Which color fills the square?",
  "options": {"A": "red", "B": "blue"},
  "images": ["data:image/png;base64,iVBORw0KGgo..."]
}

Supported image types are PNG, JPEG, GIF, and WebP. Each decoded image may be at most 20 MiB. The gateway sends the structured conversation to vLLM's /v1/chat/completions endpoint with thinking disabled, continues the exact JSON prefix, and constrains generation to one option-label token.

The pinned FP8 revision was validated end to end on one H100 80 GB: a red square selected A with 99.933% probability, a blue square selected B with 99.914%, and the text-only regression request remained correct. H200 is not required for this serving configuration.

Response:

{
  "answer": "A",
  "output": {"answer": "A"},
  "probabilities": {"A": 0.998, "B": 0.002},
  "logprobs": {"A": -0.01, "B": -6.22},
  "input_tokens": 143
}

Run the installed smoke client against a live endpoint:

mo-smoke --base-url http://localhost:8080
mo-smoke --base-url http://localhost:8080 --image ./example.png