Pattern: Open-Weights Models Close to Three Points of Frontier — The Gap Is Now a Rounding Error

OfficeChai·high signal

Kimi K2.6 at 54 vs the closed-model trio at 57 on Artificial Analysis Intelligence Index marks the smallest frontier gap ever measured for open-weights. Combined with Qwen 3.6's strong local performance (covered last week) and Gemma 4's efficient MoE architecture, open-weights models are no longer a tier below — they're a pricing decision. For builders, this means self-hosted agentic pipelines can now achieve near-frontier quality on reasoning and tool use, making the build-vs-buy calculus shift toward hybrid architectures that route by cost rather than capability.