Mozilla report: Chinese open‑weight models closing four‑month gap to frontier U.S. models

Mozilla’s State of Open Source AI v1.1 reports open‑weight models—many Chinese—are narrowing capability gaps while offering much lower token/test costs; the report warns the gap “resets every release cycle.”

In this brief: 3 sections 2 min read

- Mozilla uses METR fit on models handling tasks that take experts 8–12 hours.
- The report estimates open models reach that level about four months after closed leaders.
- Mozilla notes open capability doubling faster than closed by its computation (3.9 vs 5.5 months).
- Mozilla cites vals.ai Terminal‑Bench and Artificial Analysis indices showing shifting leaderboards.
- Example: Kimi K3 and Claude results moved between index versions; OpenRouter token volumes show heavy open‑weight traffic but closed models capture most revenue.
- Report flags large cost differences—some open models test at a fraction of closed model token cost.
- Mozilla’s dataset ends Sept. 1; later index updates change rankings.
- Report repeats an asserted (unshown) claim from an NSA/CISA/FBI advisory that Moonshot may have distilled Claude outputs to train K3.
- Authors caution the open vs closed gap can reset each release cycle.

Read full analysis on tomshardware.com

Useful?