Open weights caught the frontier. The moat is leaking.

Roll the tape: DeepSeek V4 (April, MIT licence) back near the top of the open field — confirmed by Artificial Analysis, not just the press release. Qwen (Alibaba) is the most-downloaded open family on earth, and Qwen3.6 reportedly beat a 397B-parameter model at a fraction of the size. Kimi (Moonshot) on a two-month cadence: K2.6 in April, K2.7-Code in June with +21.8% on the internal code bench, per MarkTechPost. MiniMax M3 claiming frontier coding with a million-token context window. GLM (Z.ai) as a strong all-rounder. Most from Chinese labs. None of them has to beat the latest GPT or Claude — they just have to be good enough that self-hosting becomes the rational choice for the median task.

Let's be precise about the damage. It's pricing power on the mid-tier that's at risk, not the closed labs' existence. When a capable open model exists, the incumbent can't charge a monopoly rent for an average coding task or a document summary — only for the genuine frontier edge: the hardest multi-step reasoning, the longest autonomous agent runs, the things that still require the very best models. So the closed labs get pushed upmarket, forced to actually stay ahead rather than just be exclusive. That's a worse business than 'we own all inference at every quality level'.

The comparison that makes this concrete:

For enough of the use-case distribution, row four is now the decisive one — and it tips towards open.

Here's the kicker the benchmark obsessives skip: the open floor is being set largely in Hangzhou, Beijing and Shanghai, not in San Francisco or London. DeepSeek, Alibaba/Qwen, Moonshot AI, MiniMax, Z.ai — the world's cheapest capable AI increasingly ships from Chinese labs. For most developers, the origin is secondary to the capability-cost ratio. For enterprise procurement and government operators, 'made in China' triggers supply-chain and data-trust questions a benchmark number doesn't resolve. Thunder Compute's June survey flagged it as a structural shift in the open-source centre of gravity.

However the model race ends, the open-weight wave already changed who gets to build. A developer in Jakarta or Lagos or Warsaw can now self-host a near-frontier coding model for free. That's not reversible. The question for the closed labs isn't whether to compete on the open tier — they can't win it at the margin — but whether the hard frontier edge is a defensible premium market. So far in 2026, it is. The gap is measured in months, not years, and the closed labs are feeling it. Watch whether Moonshot holds its two-month Kimi cadence, and whether MiniMax M3's weights deliver on the June launch claims once independently tested: those two data points will tell you how fast the floor rises next.