Quyet-1.2-Large
Quyet-1.2-Large is a decision model: given a state (any text, JSON or conversation) and one or more typed questions, it returns temperature-scaled probabilities over the supplied options. Question types: choice (pick one option), score (an ordered scale) and noul (true / false). It is part of the Quyet 1.2 family (Large, Small, Tiny), released by Chinh Nguyen under Apache-2.0.
| Base model | google/gemma-4-31B-it |
| Architecture | Gemma 4 31B with fully merged weights + separate option head |
| Parameters | 31.3B (30.7B text), including the option head |
| Languages | Multilingual |
| Input | 32,768 complete tokens; no truncation |
| Weights | 62.5 GB (BF16) + small F32 option head, self-contained |
| License | Apache-2.0 (see LICENSE and NOTICE) |
| Homepage | quyet.ai |
New in 1.2
- Up to 256 options per choice question; 1.0 allowed 10.
- Inputs up to 32,768 tokens, never truncated; an input over the limit is rejected instead of being cut.
- Opt-in shared prefix caching: when several questions share one state, the shared prefix can be processed once.
- The model's files are verified when it loads, and every weight hash can be checked on request.
How to use
pip install "quyet[multi-gpu]==1.2.0"
import quyet
m = quyet.load("chinhnc/Quyet-1.2-Large") # 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)}, ...}
The model folder contains the fully merged weights in the standard Gemma layout, a separate F32 option head and its tokenizer; no separate base download is needed. The package loads BF16 weights and scores the supplied options without generating text: choice uses the option head, while score and noul use vocabulary-symbol scores. It builds the input and applies the configured temperatures. Prefix caching is opt-in with prefix_cache=True and requires transformers 5.17.0. The multi-gpu extra supplies the loader dependency accelerate.
Runs in BF16 on one 80 GB GPU. In the H100 check it used about 62 GiB of allocated GPU memory, with about 76 GiB reserved, at the default batch limits, so little headroom remains; when a batch does not fit, the package splits it and retries.
Answers contain choice (with confidence and probabilities), score (expected level, confidence, probabilities, legend) or noul (P(true), with confidence). Each choice question accepts 2 to 256 options, score 2 to 10 levels, and noul exactly 2 options. Inputs are never truncated; inputs exceeding the complete-token limit are rejected.
Not evaluated for safety, bias or adversarial robustness; do not use as the only control for high-stakes decisions. In half precision, results can differ slightly between hardware and batch sizes on near-tie questions.
Credits
- Gemma 4 by Google (Apache-2.0): https://huggingface.co/google/gemma-4-31B-it.
Citation
@misc{quyet2026,
title = {Quyet 1.2: calibrated decision models},
author = {Chinh Nguyen},
year = {2026},
url = {https://huggingface.co/chinhnc/Quyet-1.2-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@quyet.ai. Package source: ncchinh/quyet.