Intern-Decision

Checkpoints	Inputs
internlm/Intern-Decision-4B, -2B, -0.8B	text

Intern-Decision models are decision models on Qwen3.5. Flide asks them with the release's own prompt. One chat turn holds the state and a schema that lists every question with its options marked A, B, C and so on, and the reply is a JSON object with one <decision> marker per question. Each answer is the softmax over its options' symbols where its marker begins, at the temperature the checkpoint's inference.py sets. Flide computes the state and the schema once and reads every question from them.

Start a server

flide serve --model internlm/Intern-Decision-4B

Send a request

curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}'

Limits and options

A request takes 1 to 16 questions. A choice takes up to 62 options, marked A to Z, a to z and 0 to 9, and a score up to 10 levels.

Option	What it does
--format-arg temperature=T	another temperature in place of the release's
--format-arg max_length=N	longest prompt in tokens, 262144 by default