Intern-Decision-4B

Use it to generate probabilistic answers to structured decision questions, including questions grounded in images.

Availability

Model weights linked in the Intern-Decision collection

Pricing model

Not established

Deployment

Local inference instructions provided

Privacy

Not established

Licence

Original Qwen license is preserved; retain the license and applicable upstream notices when redistributing.

• The weights are modified by decision tuning, and redistribution requires retaining the original Qwen license and applicable upstream notices. ^[1]
• It returns an answer distribution for every question in one model forward pass. ^[1]

  and returns an answer distribution for every question in one model forward pass.

• The page links a demo and model weights. ^[1]

  The response follows the Jev envelope: `model`, `answers`, and `usage`.

• The documented inference path runs locally using Hugging Face. ^[1]
• A request can contain 1–16 questions and up to 62 options per question. ^[1]

  Supply 1–16 questions, with up to 62 options per question.

• The model accepts a shared state, a schema of named questions, and optional images. ^[1]

  It accepts a shared state, a schema of named questions, and optional images,

• Supported question types are choice, score, and yes/no (noul). ^[2]

  `choice`, `score`, and `noul` (yes/no).

• Intern-Decision-4B is reported at 90.02% average accuracy across seven suites. ^[2]

  Intern-Decision-4B reaches **90.02% average accuracy** across seven suites (Jev: 88.74%),

• The repository provides instructions for local inference. ^[2]
• Intern-Decision takes a state, optional images, and multiple typed questions and returns decisions with probabilities. ^[2]

  Intern-Decision turns a state, optional images, and multiple typed questions into

• GitHub adds controls for Copilot code review and local sandboxing Daily brief 2026-09-26