torchcast

Checkpoints	Inputs
torchcast-ai/torchcast-decision-27b	text

torchcast is a decision model on Qwen3.5. Flide asks it with the release's own prompt. Each question is a chat prompt of its own with the state as evidence, the question and its options lettered A) to Z), and its answer is the softmax over those letters at the first token of the reply, at the temperature the checkpoint sets for the question's type. The evidence comes first in every prompt, so Flide computes it once and forks it into the questions. Thinking is off.

A choice with more than 26 options is read in rounds, as the release reads it. The first round asks the options in groups of up to 25, the next asks the top 3 of every group together, and the options left out share a small part of the probability in proportion to their first round.

Start a server

flide serve --model torchcast-ai/torchcast-decision-27b

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 choice takes up to 255 options and a score up to 10 levels. A choice with one option is answered without the model.

Option	What it does
--format-arg max_length=N	longest prompt in tokens, 262144 by default