Jev vs Mercury Decide

The analysis

Where Jev and Mercury Decide really differ, and why the bars above land where they do.

Inception released Mercury Decide on 30 September 2026 as a decision model rather than a chat model. It accepts the same request Jev does and returns probabilities read from the model’s own distribution instead of generated text, which is why output tokens cost nothing. Benchmark Heaven measured it on 7 October and added it to the board on 8 October, where it placed #3 of 135 at 72.4 against Jev’s #5 at 71.5 — its first independent score, and above Jev on all three of the benchmark’s published orders: third on the official score, fourth under B and third under C.

The composite hides a clean split. Mercury Decide is the more accurate of the two on the answers — intelligence 65.9 against 63.6, and 64.7 against 61.6 on the sealed half nobody could tune for — and the cheaper to run, 62.1 against 54.7 on the cost axis. Jev is the better calibrated by a wide margin, 90.6 against 81.3, and the faster, 91.5 against 85.5 from a 0.24s median against 0.33s. On the Capability ranking, which averages intelligence and calibration and so drops speed and cost, the order reverses: Jev is fifth at 77.1 and Mercury Decide ninth at 73.6.

We also replayed every recorded Jev request on this site — 8 homepage demos and 33 use-case templates, 41 requests and 133 questions — on Mercury Decide and compared the answers. 91.7% agree: 53 of 54 choices, 37 of 42 yes/no answers on the same side of 0.5, and 32 of 37 scores at the same level.

Agreement is not accuracy. These examples have no ground-truth labels, so where the two differ we cannot say which is right; read the numbers as how interchangeable the models are on everyday triage, routing, moderation and matching tasks. On that measure they are close, with most disagreement on borderline yes/no and graded questions rather than on clear-cut choices.

What Mercury Decide actually is

Mercury Decide, Inception’s diffusion-based decision model, by Inception.

Mercury Decide comes from Inception, the company behind the Mercury diffusion language models. Like Mercury 2.5, it refines many positions in parallel over a few passes instead of producing one token at a time. For a decision model that matters less for the answer itself than for speed: the answer is a probability over a short set of labels, not a paragraph.

It supports the three System One question types: a choice among named options, a score on an ordered rubric, and a yes/no probability. Inception aims it at moderation, routing, fraud triage, agent tool selection, evaluation loops, grading, policy checks and escalation gates — the same territory as Jev.

It is served on OpenRouter as an early-access endpoint with a 32,768-token context window. Inception notes that real latency depends on request size, network conditions, provider load and concurrency, and that a public benchmark cannot establish accuracy or calibration on your own data. Benchmark Heaven measured it through that endpoint on 7 October.

How we ran the head-to-head

Our own replay used the site’s published, fictional examples: each homepage demo and each use-case template, exactly as the playground sends them. Jev’s side is the recorded jev-1.13.0 answer already shown on those pages; Mercury’s side was run on 1 October 2026 against snapshot inception/mercury-decide-20260930, one request at a time. No request failed.

A choice agrees when both pick the same option. A yes/no answer agrees when both probabilities fall on the same side of 0.5; the mean gap between the two probabilities was 0.147. A score agrees when both round to the same rubric level; the mean gap was 0.148 of a level. The script is in the repository as scripts/bench-mercury-vs-jev.ts, so the run can be repeated.

Mercury’s median wall-clock time in this run was 0.67s, measured from one client and including the network hop to OpenRouter, so it is not comparable with lab latency. On the eight demos where Jev’s usage was recorded, Mercury counted 3,364 input tokens against Jev’s 3,749 for the same requests; the models tokenise differently.

Jev vs Mercury Decide FAQ

Other Jev comparisons

Every number on these pages is quoted from a published source and was read on October 9, 2026.

• Jev vs Imajev-4BA 4B model that reads photos, against the model it copies on text.
• Jev vs decider-4bLevel on speed, far behind on the answers.
• Jev vs JevK5A fast open Qwen3.5-4B system, and far behind Jev on the answers.
• Jev vs CygnetNo training at all, thirteenth on the board, and behind Jev on every accuracy axis.
• Jev vs decisioFourth on the board on its biggest base, and it trains nothing of its own.
• Jev vs WinnowOne local model for decisions, chat and images, against a hosted decision model.
• Jev vs djevLevel with Jev when it was last measured, and more expensive to run.
• Jev vs SemIfA fast open rebuild on your own GPU, and where it still falls short.
• Jev vs OpenJevOne name, ten projects. Here is every one, and how the drop-in compares.
• Jev vs LayaA model you fine-tune on your data, against one that works out of the box.
• Run Jev locally?The open alternatives in one table, with licences and hardware.
• Jev vs LLMHow a decision model differs from a chat LLM and a BERT classifier.

Sources

• Benchmark Heaven — JevBench v1.6.1
• Mercury Decide on OpenRouter
• Inception — Mercury Decide announcement
• TypeSafe documentation (Jev model card)