Jeffan

Checkpoints	Inputs
tattrongvu/Jeffan-4B-2010	text

Jeffan-4B-2010 is a decision model on Qwen3.5-4B. Flide asks it with the release's own prompt. Each question is a prompt of its own in the model's chat template with thinking off: a system line that tells the model to treat the state as data and reply with the code of the best option, then the state, the question and its options coded A, B, ... in the request's order, and a closing line. The answer is the softmax over the codes of the options shown at the first position of the reply. The prompt is joined from separately tokenized pieces as the release's renderer joins them, so the tokens are the release's own.

The full-attention layers read each prompt whole in both directions and the Gated DeltaNet layers stay causal, as the checkpoint's dm_config.json says, so every question is a sequence of its own and is read once.

Speed

On one H200, one request at a time, Flide answers 1.9 to 5.3 times faster than the release's own jeffan_engine on transformers 5.17. Each workload is 32 requests, each run once on both sides before timing, and the table gives the median latency of a request. MultiRC requests ask about 49 questions each.

Workload	Flide	Release	Speedup
Decision Index	15.2 ms	47.4 ms	3.1×
RACE	8.5 ms	45.5 ms	5.3×
MultiRC	305.4 ms	801.2 ms	2.6×
QuALITY long	70.9 ms	132.4 ms	1.9×

Start a server

flide serve --model tattrongvu/Jeffan-4B-2010

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 choice, noul and score questions. A choice takes up to 255 options and a score up to 10 levels. The longest prompt is 65536 tokens; a longer one is refused. A request with images is refused.

Option	What it does
--format-arg max_len=N	longest prompt in tokens, 65536 by default