Jev vs Perplexity Decider v1.1
Open Qwen3.8-27B decision model with a noncausal readout.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
Perplexity Decider v1.1
Perplexity · Open Qwen3.8-27B decision model with a noncausal readout
82.8Measured · outside the cost cap
Capability 82.8, measured and left outside the ranking cap on cost.
Key differences
- Where it runs JevHosted APIPerplexity Decider v1.1Open weights on your own GPU; the card also shows an API path
- Licence JevProprietary, hostedPerplexity Decider v1.1Apache-2.0
- Capability (v1.6.1) Jev77.1 · reference, inside the capPerplexity Decider v1.182.8 · measured, not ranked (cost)
- Intelligence / calibration Jev63.6 / 90.6Perplexity Decider v1.175.7 / 89.9
- JevBench public hard (Perplexity’s card) Jev73.27%Perplexity Decider v1.170.30%
- Estimated cost Jev$0.032 per 1,000 decisionsPerplexity Decider v1.1$0.217 per 1,000 decisions
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 $0.217 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 Perplexity Decider v1.1 is
pplx-decider-v1-27b is Perplexity’s decision model, fine-tuned from Qwen3.8-27B. It takes a state and typed questions and returns a selected choice with calibrated probabilities, in the same shape as a System One call. Benchmark Heaven measured it as Perplexity Decider v1.1 27B with a noncausal decision readout.
On JevBench v1.6.1 its Capability is 82.8, above Jev’s 77.1, with intelligence 75.7 against 63.6. Calibration is close (89.9 vs 90.6). The row is not ranked: estimated cost is $0.217 per 1,000 decisions, outside the board’s cost cap. On Perplexity’s own 11-benchmark card the Decider’s overall is 85.71% against a Jev column of 84.51%, while Jev leads the public-hard slice of JevBench on that same card (73.27% vs 70.30%).
Read this before you switch
- Outside the v1.6.1 ranking because of cost, not because it was unmeasured.
- Perplexity’s overall lead and Jev’s public-hard lead are both on Perplexity’s card. They are not a substitute for the independent Capability number.
- A 27B checkpoint is a different operations problem from a hosted call at $0.032 per 1,000 decisions.
When to use which
Choose Jev if
- You want the decision inside the cost cap, at the published $0.032 estimate.
- Public-hard items are the slice you care about — Jev leads that column on Perplexity’s own card.
- You would rather call an API than serve a 27B model.
Choose Perplexity Decider v1.1 if
- You want the higher independent intelligence score (75.7 vs 63.6) and can pay the measured cost.
- You want Apache-2.0 weights and a noncausal readout you can run yourself.
- Your eval looks like Perplexity’s broader 11-benchmark mix, where their overall is 85.71% vs 84.51%.
FAQ
Is Perplexity Decider better than Jev?
On JevBench v1.6.1 Capability, the Decider is higher: 82.8 vs 77.1, with intelligence 75.7 vs 63.6. Jev is slightly better calibrated (90.6 vs 89.9) and much cheaper on the estimate ($0.032 vs $0.217 per 1,000 decisions), which is why the Decider is measured but unranked.
Why is it missing from the ranked list?
Benchmark Heaven ranks Capability only inside a cost and median-latency cap. This row’s estimated cost is $0.217 per 1,000 decisions, so it stays in the measured-but-ineligible set.
What does Perplexity’s own table say?
On 11 benchmarks they publish, pplx-decider-v1-27b averages 85.71% and the Jev column averages 84.51%. On JevBench public hard in that same table, Jev is ahead: 73.27% vs 70.30%.