Wald

Checkpoints	Inputs
org2ai/Wald-26B-A4B	text

Wald-26B-A4B is a decision model on Gemma 4 26B-A4B, a mixture of experts with 128 experts and 8 active per token. Flide reads it as the release's own reader does, reference/eval/systemone_vllm.py in the setting its serving.json and run.sh give it.

The request is flattened as the release renders it. A string, number or boolean stays as it is, an object becomes "key: value" lines with nested values indented below their key, and an array becomes "- item" lines. A choice option is its name, or its name and description; a noul's options are "no" and "yes" with the descriptions the request gives them; a score's options are its levels. Every question is its own chat prompt with thinking off: a fixed system line, then one user turn with the state, the question, the options lettered A. to Z. and "Answer with the letter of one option."

A letter's logit combines its one-token forms, the letter alone, with a space before it and after "(", and the answer is the softmax over the question's letters. A choice of more than 26 options is read by knockout. Its options go into chunks of at most 26, each chunk is read alone, a last read picks among the chunk winners, and an option's probability is its chunk's final probability times its own within the chunk. A noul's probability of yes between 0.19 and 0.81 moves to 0.81 or 0.19 on its side, as the release's decisive floor does. The state comes first, so the questions fork from it.

Speed

On one H200, one request at a time, Flide answers 1.5 to 12.7 times faster than the release's reader on vLLM 0.30.0 with the flags of its run.sh. 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	22.2 ms	34.6 ms	1.6×
RACE	17.7 ms	32.7 ms	1.8×
MultiRC	80.8 ms	1026.8 ms	12.7×
QuALITY long	75.2 ms	115.9 ms	1.5×

Start a server

flide serve --model org2ai/Wald-26B-A4B

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 26 levels. The longest prompt is 131072 tokens, and a longer prompt is refused. A request with images is refused.

Option	What it does
--format-arg noul_floor=F	the decisive floor of a noul, 0.81 by default, 0 to turn it off
--format-arg max_model_len=N	the longest prompt, 131072 by default