Quyet-1.0-Large

Quyet-1.0-Large is a decision model: given a state (any text, JSON or conversation) and one or more typed questions, it picks one option per question and returns calibrated probabilities. Question types: choice (pick one label), score (an ordered scale) and noul (true / false). It is part of the Quyet 1.0 family (Large, Medium, Small, Small-EN, Tiny), released by Chinh Nguyen under Apache-2.0.

Base model	google/gemma-4-31B-it
Architecture	Gemma-4-31B-it with a merged LoRA fine-tune (rank 16), letter-readout decision prompt
Parameters	31.3B (30.7B text)
Languages	English; also tuned for Vietnamese. Other languages work, with lower accuracy.
Input	state up to 6,000 tokens inside an 8,000-token prompt
Weights	62.5 GB (bf16)
License	Apache-2.0 (see LICENSE and NOTICE)

How to use

pip install quyet

import quyet

m = quyet.load("chinhnc/Quyet-1.0-Large")          # pip install quyet; downloads from Hugging Face
r = m.predict(
    {"message": "Please close my card, I lost it yesterday."},
    {"intent": {"type": "choice", "instructions": "What does the customer want?",
                 "criteria": {"cancel": "close the card", "limit": "change the limit", "other": None}},
     "urgent": {"type": "noul", "instructions": "The request is urgent."},
     "mood": {"type": "score", "instructions": "How upset is the customer?", "criteria": ["calm", "annoyed", "angry"]}},
)
print(r["answers"])   # {"intent": {"choice": ..., "confidence": ..., "probabilities": {...}}, "urgent": {"noul": P(true)}, ...}

Runs in bf16 on one 80 GB GPU, or several GPUs with device_map="auto" (pip install quyet[multi-gpu]). The model answers by reading the next-token probabilities of the option letters (A, B, ...) after a fixed prompt; the quyet package builds that prompt and applies the calibrated temperatures. It also loads with transformers (and serves with vLLM) as a standard gemma-4-31B-it-architecture checkpoint, but the decision prompt and temperatures live in the package and in quyet_config.json. This model uses prompt version 2: the trained prompt without the system message and the closing line, with structured states as compact JSON (about 68 fewer input tokens per decision); its temperatures were refit for it.

Answers follow the TypeSafe /v1/systemone shape: choice (with probabilities), score (expected level, probabilities, legend) and noul (P(true)). At most 10 options per question. Only the state is ever truncated: conversation lists keep their most recent turns, other states keep their beginning.

Credits

• Gemma 4 by Google (Apache-2.0).

Citation

@misc{quyet2026,
  title  = {Quyet 1.0: calibrated decision models},
  author = {Chinh Nguyen},
  year   = {2026},
  url    = {https://huggingface.co/chinhnc/Quyet-1.0-Large}
}

License

Apache-2.0. Keep the NOTICE file (it starts with "Quyet by Chinh Nguyen") when you redistribute this model or anything derived from it. Questions and issues: email@chinh.com.