The current balance of power in open models

This is a state-of-the-union briefing on open-weight and open-source language models and the U.S.-China competitive landscape. It defines open-weight models (public weights and inference code) versus fully open-source models (which also release training code and data), and situates closed/API models on the same openness spectrum. Empirical metrics show Chinese open-weight models leading since mid-2025: Hugging Face downloads favor Chinese models by about 1.6 billion (3.2B total, roughly double U.S. totals), and benchmarks like the Artificial Analysis Intelligence Index place Z.ai’s GLM-5.3 and Moonshot AI’s Kimi K3 far ahead of leading American open models (scores in the mid-40s vs mid-20s). Chinese open-weight labs are 2-5 months behind the closed American frontier and focused on fast releases and narrower task distributions, which helps public-benchmark performance; distillation from closed APIs only slightly narrows gaps.

Adoption and risk dynamics reinforce the strategic concern: usage data from OpenRouter and other platforms show Chinese models now handling over 80% of open-model inference, and many startups and major companies (legal, coding, consumer apps) run on Chinese weights. Academia’s arXiv citations shifted similarly - open-model mentions rose from 2% in 2023 to ~50% in 2026, with Qwen surpassing Llama. Because open weights are hard to contain and restricting them would mainly harm U.S. businesses, the recommended policy response is to bolster U.S. investment and ownership of open models so domestic actors can coordinate on risk mitigation while preserving diffusion and competitiveness.