Matthew Berman examines Kimi K3 through reported frontend-development and web-engineering benchmarks, alongside a hands-on Rubik's Cube test. He distinguishes strong results on particular coding tasks from a claim that one model leads every category, and notes that the model's size makes ordinary home deployment impractical.
Matthew Berman argues that lower per-token pricing does not necessarily mean cheaper completed work. His discussion compares task completion and token consumption, emphasizing that a model requiring more reasoning tokens can lose its apparent price advantage. Reported writing and coding results are treated as source claims requiring production testing, not universal guarantees.
Matthew Berman considers how openly available models can increase competition and benefit application developers, while raising questions about infrastructure dependence and regulation. His estimates about unreleased closed models are explicitly speculative. The final demonstration shows his Rubik's Cube prototype scrambling and solving successfully, after a longer-than-expected generation run.
Watch on YouTube
