KnowLine

Checkpoints	Inputs
PelaAI/KnowLine-35B-A3B-Gen1.5	text

KnowLine is PelaAI's family of decision models. KnowLine-35B-A3B-Gen1.5 is a LoRA fine-tune of Qwen3.5-35B-A3B merged into its weights, a mixture of experts with 256 experts and 8 active per token, which Flide runs with its own grouped expert kernels. Flide reads it as the release's own front end does, knowline_server.py with knowline_layout.py, in its default setting.

The state becomes chat turns. A list of role messages, or an object with messages, stays a chat, and anything else is one user message, flattened to indented text. One more user turn holds the instruction, the question and its options labeled A, B and so on, the reply opens with Answer:, and the answer is the softmax of the option labels' logits there at temperature 1. A request whose prompt, read once, comes to at most 3000 tokens is written twice, the turns again after "Read the same input again before answering:" and the question once more, as Gen1.5 reads it. Everything up to the first question is the same for every question of a request, so the questions fork from it.

Speed

On one H200, one request at a time, Flide answers 1.2 to 5.4 times faster than the release's knowline_engine.py on SGLang 0.5.21 with the flags of its serve_knowline.sh, which loads the weights in FP8 and keeps the output head in float32. Flide runs with --quantize fp8, which puts the routed experts in FP8 and keeps the rest of each layer in 16 bits. 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	32.2 ms	41.2 ms	1.3×
RACE	23.3 ms	40.2 ms	1.7×
MultiRC	69.3 ms	377.3 ms	5.4×
QuALITY long	64.7 ms	80.6 ms	1.2×

Start a server

flide serve --model PelaAI/KnowLine-35B-A3B-Gen1.5 --quantize fp8

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, up to 64 of them, and a question up to 255 options or levels. A request with images is refused.

Option	What it does
--quantize fp8	the routed experts in FP8, as the speeds above are measured
--format-arg read_twice=0	every request read once
--format-arg twice_max_tokens=N	the length up to which a request is read twice, 3000 by default
--format-arg temperature=T	the temperature of the answers, 1 by default