Quyet-1.0-Large's homepage prints JevBench v1.6.0 at 71.39

Florian S posted on October 5, 2026, at 15:09 UTC, that JevBench v1.6.0 was live. Capability rank 1 in that post is Quyet-1.0-Large at 81.7, ahead of deck-31B at 77.6 and Jev 1.13.0 at 76.5. The post says both models ahead of Jev are Gemma 4 checkpoints with open weights.

The homepage, opened October 12, prints a frozen v1.6.0 table and links benchmarkheaven.com/jev-models/v1.6.0. We did not reopen that file. The table says the release scored 1,624 decisions, open and sealed, across 92 systems.

System	Official score	Rank	Capability	Intelligence	Calibration	$ / 1,000
Quyet-1.0-Large	71.39	1	81.7	73.4	90.0	0.045
Jev 1.13	71.09	2	76.5	62.4	90.6	0.032
Clef	16.75	43	75.3	59.1	91.6	0.249
decider-35b-a3b	24.77	29	66.2	48.7	83.6	0.154
decider-4b	41.24	17	64.9	40.1	89.6	0.015

The homepage says Capability is the mean of Intelligence and Calibration, ranked among systems at most twice Jev’s cost and latency. deck-31B at 77.6 is the October 5 post. It is not a row in the homepage table we copied.

Later readings on this site stay on their own revisions. The October 8 open-weights list still leads with Capability 81.7, and the rank line says official composite 3. H2O’s October 7 blog prints composite 71.4 for Quyet-1.0-Large on revision v1.7.17, behind H2O-Lightning-4B at 72.5 and the Jev reference at 71.5. v1.6.1 on the JevBench page scores 1,500 decisions. Jev’s capability on that later board is 77.1. The frozen table’s 76.5 is the v1.6.0 cell.

Microsoft-Decision-1 says it was 4.5 times quicker than Quyet-1.0-Large. That post does not print an accuracy, and the appendix does not name Jev. The speed line is Microsoft’s comparison.

Quyet is Chinh Nguyen’s family of decision models. A state and one or more typed questions come back as a probability over the options the caller supplied. The README says the answers follow the TypeSafe /v1/systemone shape: choice, score, and noul. The package is pip install quyet. License on the Large card is Apache-2.0, and the card says a NOTICE file has to stay with a redistribution. The text is tuned for English and Vietnamese. Small-EN is English only.

Checkpoint	README parameters	Homepage line
Large	31.3B, 30.7B text	Gemma 4 31B, 80 GB GPU
Medium	4.66B, 4.21B text	Qwen3.5 4B, 4.7B, 16 GB GPU
Small	328M	Encoder, 330M
Small-EN	153M	Encoder, 150M
Tiny	183M	Encoder, 180M, 6 ms

The Large card says the checkpoint is Gemma-4-31B-it with a merged LoRA of rank 16, read out from the next-token probabilities of the option letters. Weights are 62.5 GB in bf16. The state is capped at 6,000 tokens inside an 8,000-token prompt. A question takes at most 10 options. Prompt version 2 drops a system message and a closing line, and the card says that cuts about 68 input tokens. Temperatures were refit for that prompt. The card says one 80 GB GPU runs bf16, or several GPUs with device_map="auto".

The Small card says the encoder is aisingapore/SEA-LION-ModernBERT-300M plus a two-layer head, 328M parameters, input 8,192 tokens, 1.31 GB in F32, Apache-2.0. The homepage’s 330M is the rounded cell.

The homepage names a hosted listing of Large at FPT AI. That page was not opened, and it is not a row in the catalog here. The homepage’s demo box was not submitted. We did not run the weights.