Qwen3.5

Checkpoints	Inputs
Qwen/Qwen3.5-0.8B, -2B, -4B, -9B, -27B	text, images

Any Qwen3.5 chat model serves as it is. Each question is asked in the user turn after the state, with one line per option labeled A, B, C and so on and a line that asks for one label, and its answer is read from the probabilities of those labels at the first token of the reply. A noul is answered yes or no, and score levels are labeled 0 to 9, or with letters past ten levels. After Z come two-letter labels, AA, AB and so on, each one token, so a question takes up to 256 options or levels. Thinking is off.

Start a server

flide serve --model Qwen/Qwen3.5-4B

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?"}
  }
}'

Images

Images go in images, as data URLs or plain base64, and the model sees them ahead of the text.

IMG=$(base64 -w0 photo.jpg)
curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d @- <<EOF
{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund. A photo is attached.",
  "images": ["data:image/jpeg;base64,$IMG"],
  "questions": {
    "cracked": {"type": "noul", "instructions": "Does the photo show a cracked screen?"},
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}}
  }
}
EOF

Images need torchvision, installed from the same place as torch.

Options

Option	What it does
--quantize fp8	decoder weights in FP8 on Ada, Hopper and Blackwell GPUs; the numbers move a little
--max-branch-tokens 32768	longest state plus question, in tokens
--max-request-tokens 65536	longest request, the state once plus every question, in tokens
--format-arg style=flide	the prompt Flide used by default before 0.1.2; text only
--format-arg style=sglang	the prompt SGLang's /v1/systemone renders for a chat model

flide serve lists the options every family shares.