Quyet models read any state (text, JSON or a conversation) and answer typed questions with one option and a calibrated probability, in one forward pass. Open weights, Apache-2.0.
Pick an example, edit the state, and see how the calibrated answers change.
Protected by Cloudflare Turnstile and a per-visitor rate limit. Demo inputs are not stored.
Your code asks a typed question and gets back a typed answer with a probability. Quyet never generates text, so there's no parsing and no off-list answers.
pip install quyet
import quyet
m = quyet.load("chinhnc/Quyet-1.0-Small") # CPU is fine for Small / Tiny
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"}}})
print(r["answers"]["intent"]) # {"choice": "cancel", "confidence": 0.99, "probabilities": {...}}
Same questions and answer format everywhere. Pick by accuracy, speed and where it runs.
| Model | Type | Params | Runs on | Weights |
|---|---|---|---|---|
| Quyet-1.0-Large | LLM (Gemma 4 31B) | 31.3B | 80 GB GPU | Hugging Face |
| Quyet-1.0-Medium | LLM (Qwen3.5 4B) | 4.7B | 16 GB GPU | Hugging Face |
| Quyet-1.0-Small | Encoder | 330M | CPU or any GPU | Hugging Face |
| Quyet-1.0-Small-EN | Encoder | 150M | CPU or any GPU | Hugging Face |
| Quyet-1.0-Tiny | Encoder | 180M | CPU or any GPU · 6 ms | Hugging Face |
Official JevBench v1.6.0 results, run independently by Benchmark Heaven on 1,624 decisions (open and sealed). Quyet-1.0-Large leads both the official ranking and the Jev-class Capability ranking.
| System | Official score | Rank | Capability | Intelligence | Calibration | Cost / 1K decisions | Weights |
|---|---|---|---|---|---|---|---|
| Quyet-1.0-Large | 71.39 ★ | #1 | 81.7 ★ | 73.4 ★ | 90.0 | $0.045 | open |
| Jev 1.13 (TypeSafe) | 71.09 | #2 | 76.5 | 62.4 | 90.6 | $0.032 | closed API |
| Clef (Cloudflare) | 16.75 | #43 | 75.3 | 59.1 | 91.6 | $0.249 | open |
| decider-35b-a3b (Mapika) | 24.77 | #29 | 66.2 | 48.7 | 83.6 | $0.154 | open |
| decider-4b (Mapika) | 41.24 | #17 | 64.9 | 40.1 | 89.6 | $0.015 | open |
Source: benchmarkheaven.com/jev-models/v1.6.0 (frozen release, 92 systems). Capability = mean of Intelligence and Calibration, ranked among Jev-class systems (at most 2× Jev's cost and latency).
Wherever software needs one answer and how sure it should be.
Quyet-1.0-Large is available from FPT AI.
Hosting Quyet as an inference provider? Email email@quyet.ai to get listed here.