StartLux¶
StartLux-Decision v1.1 Instant are StartLux's decision models on Qwen3.5, from 0.8B to 27B. Flide reads them as their
startlux-decide server reads them. Each question is a chat prompt of its own with thinking off: a fixed system line,
then "Evidence:" with the state, the question and its options lettered A) to Z) in the request's order. The answer is
the softmax over the listed letters at the first token of the reply, at the temperature that decision_config.json
sets for the question's type. Every choice gets "None of the above" as an extra last option, as the server adds it;
its share is left out and the listed options are renormalised.
The full-attention layers read each prompt whole in both directions and the Gated DeltaNet layers stay causal, as
decision_config.json says, so every question is a sequence of its own.
A choice with more than 25 options is read in rounds, as the server reads it. The first round asks the options in groups of up to 25, the final round asks the best 3 of every group (the best 25 when more qualify) with the extra option, and the options left out share a small part of the probability.
Speed¶
Flide answers 1.1 to 3.3 times faster than the startlux-decide server on vLLM 0.31.0 with its plugin, one request at a time. 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.
StartLux-Decision-v1.1-4B-Instant on one H200
| Workload | Flide | Release | Speedup |
|---|---|---|---|
| Decision Index | 13.6 ms | 35.8 ms | 2.6× |
| RACE | 8.5 ms | 27.8 ms | 3.3× |
| MultiRC | 298.8 ms | 328.3 ms | 1.1× |
| QuALITY long | 67.1 ms | 79.9 ms | 1.2× |
StartLux-Decision-v1.1-2B-Instant on one A40
| Workload | Flide | Release | Speedup |
|---|---|---|---|
| Decision Index | 27.5 ms | 64.0 ms | 2.3× |
| RACE | 17.1 ms | 54.6 ms | 3.2× |
| MultiRC | 648.8 ms | 768.5 ms | 1.2× |
| QuALITY long | 140.2 ms | 185.3 ms | 1.3× |
StartLux-Decision-v1.1-0.8B-Instant on one A40
| Workload | Flide | Release | Speedup |
|---|---|---|---|
| Decision Index | 15.7 ms | 34.1 ms | 2.2× |
| RACE | 9.5 ms | 27.8 ms | 2.9× |
| MultiRC | 348.2 ms | 402.7 ms | 1.2× |
| QuALITY long | 84.7 ms | 98.3 ms | 1.2× |
Start a server¶
flide serve --model startlux-models/StartLux-Decision-v1.1-4B-Instant
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. A choice that names one of its options "None of the above" is refused, since the server adds that option itself. The longest prompt is 131072 tokens, the server's context for the Instant models. A request with images is refused.
| Option | What it does |
|---|---|
--format-arg max_length=N |
longest prompt in tokens, 131072 by default |