JevAny

Checkpoints	Inputs
SimpleJev/JevAny-Qwen3.5-4B-LoRA	text, images

JevAny is SimpleJev's family of decision models with a pointer readout, here a LoRA adapter on Qwen3.5-4B. Flide merges the adapter into the base weights as they are read, in float32 and rounded once, as its release serves it, and tokenizes with the tokenizer the checkpoint saved.

A request is laid out as for Kev. The state comes first, after <|fim_prefix|>, and every question is a branch of its own: <|fim_middle|> and its instructions, each option between <|box_start|> and <|box_end|>, then <|fim_suffix|>. A checkpoint whose tokenizer names other delimiters gets those. The pointer head in head.pt reads each question at its last token and at every option. An option's logit is the dot product of its key with the question's query, divided by 16, plus the head's residual MLP over their product, all divided by the temperature that head.pt carries, and the answer is the softmax over the options. The state is computed once and forked into the questions.

A request with images asks one question. The whole request is then one text with the images' placeholders ahead of it, read through the checkpoint's processor, and n images are resized to at most 512 × 512 / n pixels each (128 × 128 when there are more than 16), as the release sizes them.

Speed

On one H200, one request at a time, Flide answers 1.1 to 5.9 times faster than the release's JevModel with its serving settings (bf16, the adapter merged, its CUDA graphs and fused Qwen3.5 kernels). Each workload is 32 requests, each run once on both sides before timing, and the tables give the median latency of a request. MultiRC requests ask about 49 questions each, and GQA requests ask one question about an image.

JevAny-Qwen3.5-4B-LoRA

Workload	Flide	Release	Speedup
Decision Index	10.6 ms	15.7 ms	1.5×
RACE	8.6 ms	9.5 ms	1.1×
MultiRC	60.5 ms	354.0 ms	5.9×
QuALITY long	61.9 ms	72.7 ms	1.2×
GQA	18.4 ms	80.9 ms	4.4×

Start a server

flide serve --model SimpleJev/JevAny-Qwen3.5-4B-LoRA

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 up to 64 choice, noul and score questions. A choice takes up to 256 options and a score from 2 to 255 levels, and a noul's criteria name only false and true. The state, a question's row (the state and its branch) and the whole request each take up to 8,192 tokens, and a longer one is refused, as the release refuses it. A request may carry up to 32 images (data URLs or base64) with one question.

Option	What it does
--format-arg temperature=T	the temperature of the answers, the one head.pt carries by default