Quyet¶
| Checkpoints | Inputs |
|---|---|
chinhnc/Quyet-1.0-Medium |
text |
Quyet-1.0-Medium is a decision model on Qwen/Qwen3.5-4B. Flide asks it with the release's own prompt. Each question
is a chat prompt of its own. The system turn tells the model to choose exactly one option, using only the state and the
option descriptions, and to reply with the option letter. The user turn holds the state, the question, a line that
depends on the question's type, and the options lettered A. to J., and ends by asking for one letter. A choice lists each
option's description, or its name when it has none. A score lists its levels, lowest first. A noul reads "Choose A if
the statement is true for this state, B if it is not." and lists the request's descriptions of true and false, or the
words true and false, in that order.
A string state is shown as it is, and an object or a list is shown as JSON indented by one space. The state is cut to
the budget that limits.max_state_tokens in quyet_config.json sets. A list keeps its last tokens behind a leading
…, and any other state keeps its first tokens followed by …. While the longest prompt of the request is over
limits.max_prompt_tokens, the budget shrinks by the excess plus 16 tokens and the prompts are rendered again, up to
four tries in all.
The answer is the softmax over the letters A to J at the first token of the reply, each letter read as the token of the
letter alone. The temperature of each question type comes from the temperatures entry of quyet_config.json. The
state comes first in every prompt, so Flide computes it once and forks it into the questions. Thinking is off.
Speed¶
On one H200, one request at a time, Flide answers 1.4 to 7.0 times faster than Quyet's own package quyet 1.0.2 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 | 8.8 ms | 52.6 ms | 5.9× |
| RACE | 9.0 ms | 53.2 ms | 5.9× |
| MultiRC | 76.2 ms | 535.0 ms | 7.0× |
| QuALITY long | 75.0 ms | 102.4 ms | 1.4× |
Start a server¶
flide serve --model chinhnc/Quyet-1.0-Medium
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 10 options, labelled A to J, and a score up to 10
levels. The longest prompt is the limits.max_prompt_tokens of the checkpoint's quyet_config.json, and the state
takes at most limits.max_state_tokens of it. A request with images is refused.
The template refuses a question whose instructions are missing, empty or not a string, a choice or a score with fewer
than two options, and a state that is null or a number. It refuses a request whose prompt is still too long after the
fourth try, or whose state budget would fall under limits.min_state_tokens.
| Option | What it does |
|---|---|
--format template --format-arg spec=chinhnc__quyet-1.0-medium |
reads a copy of the checkpoint kept in a directory Flide does not recognize |