Open source, self-hosted. There is no hosted API. You download the weights and run them on your own hardware. How open-source models are listed.
What Quyet-1.0-Large is
Quyet-1.0-Large is a decision model released by Chinh Nguyen under Apache 2.0. It is Google’s Gemma-4-31B-it with a rank-16 LoRA fine-tune merged into the weights, 31.3 billion parameters in all. It is the largest of five Quyet 1.0 sizes, which run down to a 180-million-parameter encoder that works on a CPU.
It reads a state, which can be text, JSON or a conversation, and answers typed questions about it: Choice with up to 10 options, Score and Noul, the same three shapes Jev uses. The state can hold up to 6,000 tokens inside an 8,000-token prompt.
How it answers
It does not write text. A fixed prompt lists the options as letters, and the model reads the probability of each letter as the next token, so one forward pass gives the whole answer. The quyet Python package builds that prompt and applies temperatures fitted for calibration. The card says the checkpoint also loads in transformers and vLLM, but the prompt and temperatures live in the package and a config file. Answers follow the /v1/systemone shape, and quyet.ai says any provider that hosts the model uses that route. quyet.ai names FPT AI as one such provider.
What it is good at
Accuracy among open-weights systems. Benchmark Heaven’s JevBench ranks it first on Capability, the average of Intelligence and Calibration, at 81.7. It scores 73.4 on Intelligence and 90.0 on Calibration. On the official composite, which also weighs speed and cost, it is third at 71.4. The Jev Decision Index (multimodalart’s community board on Hugging Face) gives it a Full score of 58.71 and a public score of 60.79, with an expected calibration error of 0.0778. Jev 1.13.0 is the reference row on JevBench.
Microsoft’s launch post for Microsoft-Decision-1 scored it 81.9% accuracy and 93.1 calibration on a 36-benchmark suite Microsoft ran itself and did not name. The card is tuned for English and Vietnamese.
What it is not for
It is heavy. The weights are 62.5 GB and the card asks for one 80 GB GPU, or several smaller ones. Its median of about 105 to 116 milliseconds is slower than the 4B models on the same board, such as H2O-Lightning-4B at 29 milliseconds. A question takes at most 10 options, so large label sets need a shortlist first. Other languages work with lower accuracy.
The Jev Decision Index row carries a note saying the public run was done by the lab with the option cap lifted, and the index does not say more. We could not read the FPT AI listing, so its price and limits are unknown.
Specifications
| System One Score | 96.1 from 2 of 3 boardsJev Decision Index 93.6, JevBench 98.7. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | 10 |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 8,000 tokens |
| State budget | 6,000 tokens |
| Rate limit | None when self-hosted. |
| Licence | Apache-2.0 |
| SDKs | Python: quyet |
The state can hold up to 6,000 tokens inside an 8,000-token prompt. Only the state is truncated: conversation lists keep their most recent turns, and other states keep their beginning. Each question takes at most 10 options. The model runs in bf16 on one 80 GB GPU (62.5 GB of weights), or across several GPUs with the quyet[multi-gpu] extra. It is tuned for English and Vietnamese; other languages work with lower accuracy.
Versions
- chinhnc/Quyet-1.0-Large, 3 Oct 2026, Gemma-4-31B-it with a rank-16 LoRA merged in, 31.3B parameters. The card describes prompt version 2, with refit temperatures and about 68 fewer input tokens per decision. The same family has Medium (Qwen3.5-4B), Small, Small-EN and Tiny. Release notes
Use cases
What people use Quyet-1.0-Large for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
LLM guardrails with System One models
Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path.
Compliance verification with System One models
A compliance review is a fixed checklist run against a changing document. Jev answers the whole checklist in one request, returning a probability, a label or a rated level per item, so code can clear the clear-cut findings and send the borderline ones to a person.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.3.1 | Full score | 58.71 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 60.79 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.0778 ECE | Independent | 12 Oct 2026 |
| JevBench 1.7.46 | Composite Score | 71.4 score | Independent | 12 Oct 2026 |
| JevBench 1.7.46 | Capability Score | 81.7 score | Independent | 12 Oct 2026 |
| Microsoft 36-benchmark comparison | Average accuracy | 81.9 % | vendor-run | 9 Oct 2026 |
| Microsoft 36-benchmark comparison | Calibration score | 93.1 score | vendor-run | 9 Oct 2026 |
- Jev Decision Index, Full score: Row: Quyet-1.0-Large. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect. The row's note reads: Public board run by the lab; option cap lifted.
- Jev Decision Index, Public score: Row: Quyet-1.0-Large.
- Jev Decision Index, Expected calibration error: Row: Quyet-1.0-Large. Lower is better.
- JevBench, Composite Score: Row: Quyet-1.0-Large. Third among open-weights systems on the official composite. Intelligence 73.4, Calibration 90.0. The board re-measures rows, so this can differ from earlier editions.
- JevBench, Capability Score: Row: Quyet-1.0-Large. First in the Capability ranking, which averages Intelligence and Calibration and leaves out speed and cost.
- Microsoft 36-benchmark comparison, Average accuracy: Run by Microsoft on an undisclosed 36-benchmark suite, 147,137 questions. Microsoft-Decision-1 scored 83.5% on the same suite. The benchmark names are not given, so the figure cannot be reproduced.
- Microsoft 36-benchmark comparison, Calibration score: Run by Microsoft on an undisclosed 36-benchmark suite; 100 is perfect.