K3 Predicted to Hit Top 10 on WeirdML

Teortaxes, known as a DeepSeek booster, shared a two-week assessment of K3 under Harvard's policy barring tests of China-served models. Citing GLM progress, the prediction placed K3 between 75-83 percent across settings and inside the top 10. Florian Brand replied that 81-86 percent is plausible, though tool-use harness issues may limit results. For certain research tasks the model reaches frontier level. The short window and policy constraints frame the current confirmed projections without external verification.