Will the highest score achieved by a Moonshot Kimi model on Humanity’s Last Exam in 2026 be 50% or higher?

Analysis

The decisive question is whether tool-augmented self-reports count. On that basis Kimi needs a no-tools ≥50% from a new model within ~70 days: K3 is at 43.5-46.9% on lenient sources and absent from the official Scale board (K2.5 at 24.4%), and official runs undercut vendor numbers by 7-13 points. K3.1 is a point release teased with coding/3D demos; a 3-6.5-point no-tools jump plus timely official listing is possible but not the base case. I put P(YES) ≈ 25% (allowing for a strong K3.1 thinking variant and some resolver leniency on source).

Key Evidence

Resolvers key on official/no-tools leaderboard figures. Kimi's best no-tools figures are 43.5% (model card) / 46.9% (Artificial Analysis); the Scale/CAIS official board lists only kimi-k2.5 at 24.37% with no K3 entry, and official runs score 7-13 pts below vendor numbers. Only K3.1 (a point release) remains in the ~70-day window.

Risks

Kimi K3.1 (teased today) ships as a reasoning-heavy release whose model card reports ≥50% on HLE without tools, Artificial Analysis or the official board reproduces it by late December, and the resolver accepts it — or the resolver unexpectedly treats Moonshot's 56% with-tools figure as the 'HLE Accuracy'.