Open Models Are the United Front of AI: How Qwen 3.8 27B Changes the Meta

ChinaTalk · rss · 2026-09-06

A ChinaTalk translation of a Chinese AI blog essay analyzing the open-source vs closed-source race:

The basics: OpenAI and Anthropic still hold a clear lead; Chinese models like K3 and Qwen 3.8 Max approach them on benchmarks but remain a full generation behind, with recent progress mainly in faster follow-up speed.
Open source is the key variable: Open-weight releases have consolidated all teams outside closed labs into a de facto united front. When the US government floated restricting open-source model deployment after K3's release, over 270 companies including Nvidia and Microsoft signed an open letter in its defense — treatment a closed model would never get.
Small models change the meta: Qwen 3.8 27B hits roughly Opus-4.6-level performance with 1/28 the parameters of GLM 5.2 (744B), sparking the "intelligence density" debate. The author argues the concept doesn't survive scrutiny — the real ticket to the new season is still 2T-parameter models — but 27B makes a model surpassing human coding ability a true public good runnable on consumer GPUs.
Cabbage-cheap intelligence: Citing Xiaozhi AI (1M+ units shipped, born from spare 4090s and open-sourced Qwen 2.5), the author predicts wilder creations once frontier-level intelligence becomes trivially cheap; most future token demand will be served by open models.

More from AGI Musings