Shanghai AI Laboratory releases Intern-Decision, multimodal decision models from 0.8B to 4B
Intern-Decision-4B is fine-tuned from Qwen3.5-4B and answers up to 16 questions per pass, with image input. Apache-2.0.
Shanghai AI Laboratory published the Intern-Decision family on Hugging Face in late September 2026. The 4B repository was created on 26 September; the card gives no release date, so that is the date used here. Sizes visible on the Hub include 0.8B, 2B and 4B.
The 4B card describes a multimodal structured decision model fine-tuned from Qwen3.5-4B. It takes a shared state, named questions and optional images, and returns an answer distribution for every question in one forward pass. Limits are up to 16 questions, 62 options per question, 8 images and an 8,192-token input.
The licence is tagged Apache-2.0. The card says the original Qwen licence is kept as LICENSE-QWEN and upstream notices must be retained on redistribution. A demo Space, a weights collection and a GitHub repo are linked.
The card reports Jevbench Easy, Original and Hard at 100.00, 98.61 and 73.87, a Typed Decision score of 80.55, and an average of 90.02. It reports Brier 0.347 and ECE 0.065, and a mean latency of 44.16 ms on one RTX 4090.
Easy at 100 means the easy set is solved. The Hard number is the one to read.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights, a demo Space and GitHub are linked on the card.
- Useful 3.5: Small enough to run locally; image input is a plus over text-only peers.
- Notable 3: A major open lab entering the category.
- Novel 3: Same decision format; multimodal input is the difference.
Comments