Does the standard hold across Moonshot AI's models?

Every model the Benchmark tests from Moonshot AI, checked against one line: with guidance, fewer than 10 in 100 drafts needing revision.

Human-centered language, with guidance

68 drafts per model · bar shows the 95% interval · dashed line at 10%

Kimi K3 Settled, predecessor in a pair

2 of 68 need revision (2.9%), interval 0.8 to 10.1%. Without guidance: 32.4%.

Can’t tell yet

Kimi K3.1 Pro First look, paired

2 of 68 need revision (2.9%), interval 0.8 to 10.1%. Without guidance: 10.3%.

Can’t tell yet

Holds: the whole interval sits under the line. Falls short: the whole interval sits above it. Can't tell yet: the interval crosses the line, and more drafts are needed to say.

Trauma-informed language, with guidance

24 drafts per model · bar shows the 95% interval · dashed line at 10%

Kimi K3 Settled, predecessor in a pair

0 of 24 need revision (0.0%), interval 0.0 to 13.8%. Without guidance: 33.3%.

Can’t tell yet

Kimi K3.1 Pro First look, paired

0 of 24 need revision (0.0%), interval 0.0 to 13.8%. Without guidance: 20.8%.

Can’t tell yet

Holds: the whole interval sits under the line. Falls short: the whole interval sits above it. Can't tell yet: the interval crosses the line, and more drafts are needed to say.

Generations in this lab

Kimi K3 to Kimi K3.1 Pro, same run. With guidance: no detectable change. Without guidance: improved.

New models, in detail

Moonshot AI models the Benchmark doesn't test

This list comes from the lineage registry, which isn't built yet. It will name each untested model and why, so silence never reads as a pass.

Table design inspired by EQ-Bench, by Sam Paech. Benchmark engineered by MKTNG.