LingeeJevMulti-4B-v1

LingeeJevMulti-4B-v1 is a 4.54B multimodal typed-decision model based on Qwen3.5-4B.

It reads text and images and returns native probabilities for typed choice, noul and score decisions without generating explanatory text.

Property	Value
Parameters	4.54B, including a 0.33B vision encoder
Inputs	Text or structured state, plus one or more images
Image limit	1,638,400 pixels per image
Outputs	Choice, Yes/No and Score probability distributions
API	POST /v1/systemone
License	Apache-2.0

Quickstart

python systemone.py \
  --model . \
  --host 127.0.0.1 \
  --port 8000

curl -s http://127.0.0.1:8000/v1/systemone \
  -H 'content-type: application/json' \
  -d '{
    "state": "",
    "questions": {
      "q": {
        "type": "choice",
        "instructions": "Choose the matching color.",
        "criteria": {"A": "red", "B": "blue"}
      }
    }
  }'

Benchmark compatibility

• Native typed-decision probability API.
• Choice, Noul and Score supported.
• Images are capped at 1.6MP.
• Local 1.6MP H200 warm measurements are included in benchmark_1p6mp.json.
• Official leaderboard results are published only after the benchmark maintainers evaluate this release.

Reproducible inference

Use Python and package versions in requirements.txt. The model directory contains all required model configuration, tokenizer, processor, Decision Readout and runtime files. COMPLETE.json contains SHA-256 integrity hashes for the release files.

License

Distributed under Apache-2.0. Retain LICENSE and NOTICE when redistributing.