Jev vs Quyet-1.0-Large
Apache-2.0 Gemma-4-31B decision model, first on the open-weights Capability board.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
Quyet-1.0-Large
Chinh Nguyen · Apache-2.0 Gemma-4-31B decision model, first on the open-weights Capability board
81.7Open-weights Capability #1
Capability #1 at 81.7. Intelligence 73.4, the highest eligible score on that axis.
Key differences
- Where it runs JevHosted API, generally availableQuyet-1.0-LargeSelf-hosted, one 80 GB GPU in bf16 (62.5 GB weights)
- Licence JevProprietary, hostedQuyet-1.0-LargeApache-2.0 weights and package
- Capability (v1.6.1) Jev77.1 · unranked referenceQuyet-1.0-Large81.7 · rank #1
- Intelligence / calibration Jev63.6 / 90.6Quyet-1.0-Large73.4 / 90.0
- Input size Jev64k tokens per requestQuyet-1.0-LargeState up to 6,000 tokens inside an 8,000-token prompt
- Languages JevThe hosted English decision APIQuyet-1.0-LargeEnglish, also tuned for Vietnamese
JevBench v1.6.1, same method
Capability is the mean of intelligence and calibration, ranked only inside Benchmark Heaven’s cost and median-latency caps. Jev 1.13.0 stays on the board as the reference row and is not given a rank. A model can score above that reference and still be left off the numbered list when its estimated cost is too high. Full table ↗
Capability axes, higher is better
Cost $0.045 per 1,000 decisions (estimated). Median latency 0.4 s. Jev’s reference row is $0.032 per 1,000 decisions (not published as a list price), median 0.2 s. Both dollar figures are the benchmark’s, except Jev’s, which Benchmark Heaven marks as not a published list price.
What Quyet-1.0-Large is
Quyet-1.0-Large is Chinh Nguyen’s Apache-2.0 decision model on Gemma-4-31B-it: a merged LoRA (rank 16) that answers by reading the next-token probabilities of option letters, then applies calibrated temperatures. It speaks the same job as Jev — choice, score, and true/false — and ships in a `quyet` package that builds the prompt.
On JevBench v1.6.1 it is the open-weights Capability leader at 81.7, ahead of Jev’s 77.1 reference. The lead is intelligence (73.4 vs 63.6). Calibration is close, and Jev is the higher of the two (90.6 vs 90.0). Estimated cost is $0.045 per 1,000 decisions against Jev’s $0.032, and median latency is 0.4 s against 0.2 s.
Read this before you switch
- The prompt budget is about 8,000 tokens, with the state itself capped around 6,000 — a small fraction of Jev’s 64k-token request.
- Weights are 62.5 GB in bf16. The author documents one 80 GB GPU, or several GPUs with device_map="auto".
- Languages other than English and Vietnamese are described as working at lower accuracy.
When to use which
Choose Jev if
- The request is a long document, trace, or ticket past a few thousand tokens.
- You want a hosted endpoint and a published per-token price instead of an 80 GB GPU.
- You threshold on probabilities and want the slightly higher calibration score (90.6 vs 90.0).
Choose Quyet-1.0-Large if
- You need the highest eligible intelligence score on the October board (73.4).
- You want Apache-2.0 weights, including a Vietnamese-tuned prompt.
- You can hold a 31B model on an 80 GB GPU inside your own network.
FAQ
Is Quyet better than Jev?
On JevBench v1.6.1 Capability, yes: 81.7 against Jev’s 77.1 reference, driven by intelligence 73.4 against 63.6. Jev is slightly better calibrated (90.6 vs 90.0), cheaper on the published estimate ($0.032 vs $0.045 per 1,000 decisions), and faster at the median (0.2 s vs 0.4 s).
Is Quyet open source?
Yes. Apache-2.0 weights and the quyet package, by Chinh Nguyen. It is independent of TypeSafe. The family also includes Medium, Small, Small-EN, and Tiny checkpoints with lower Capability scores.
Can Quyet replace a Jev call?
The question types match — choice, score, and true/false — but the state has to fit in about 6,000 tokens. Longer Jev requests will not fit unchanged.
What hardware does Quyet-1.0-Large need?
The model card says bf16 on one 80 GB GPU, or several GPUs with device_map="auto". The checkpoint is 62.5 GB.