One Went Dark, Two Went Open

This week ran a controlled experiment on who’s winning AI, entirely by accident. Hold two dates next to each other.

June 12: the US government export-controlled Fable 5 off the planet, and its own biggest investor pulled the trigger. The most capable model America had ever shipped went dark on a Friday-night letter.

Also June 12, and June 13: China’s open-weight labs shipped two major coding models into the commons and walked away. Nobody can take them back.

What Actually Dropped

Kimi K2.7 Code (Moonshot, June 12) landed open on Hugging Face on day one: a trillion-parameter Mixture-of-Experts model, 32B active across 384 experts, 256K context, under a Modified MIT license. Moonshot reports +21.8% on its own Code Bench v2 over K2.6 while burning roughly 30% fewer reasoning tokens, at about $0.95 in / $4.00 out per million. Coding-first, agentic, and self-hostable.

GLM-5.2 (Zhipu, June 13) shipped across every GLM coding tier with a 1M-token context window, five times GLM-5.1’s, with MIT open weights promised within the week.

Behind the two headliners, the pack is the real signal. DeepSeek V4 Pro Max (80.6% SWE-bench Verified, MIT), MiniMax M3 (a tenth of the incumbent’s price), and Qwen3.7 Max now cluster within about 0.2 points of Gemini 3.1 Pro on verifiable coding. Kimi K2.6 was already the first open model to beat GPT-5.4 on SWE-Bench Pro. This is not a frontier-beating story. It is a “good enough, open, and a fraction of the price” story, which is the only one that moves a market.

The Diffusion Layer Didn’t Blink

I argued last week that the US is winning the trophy and ceding the territory: the frontier benchmark it photographs versus the diffusion layer that compounds. This week was the thesis run as a lab experiment, and the result printed cleanly.

While Washington and Amazon spent 72 hours staging an export-control circus around one closed model, the open-weight ecosystem shipped two more and kept moving. It did not pause for the panic. It accelerated through it, because nothing about a Friday-night letter in DC slows a release out of Beijing.

And the cadence is the quiet part. Kimi K2.7 is Moonshot’s fifth major release in under a year. GLM-5.2 follows GLM-5.1 by weeks. DeepSeek, Qwen, and MiniMax are all shipping on the same metronome. The frontier labs ship a flagship a quarter with a launch event; the open pack ships something every few weeks and lets the weights do the marketing. One of those tempos compounds faster, and it is not the one with the keynote.

That’s the sound of millions of companies and users doubling down on Chinese models.

— zmmmmm, top of the Hacker News thread on the Fable ban

The Asymmetry Is the Whole Story

Here is the structural fact the week made undeniable: you can switch off a closed model you host. You cannot recall a weights file already on ten thousand laptops.

Fable 5 went dark because Anthropic serves it, so a directive to Anthropic is a directive to every user at once. Kimi K2.7 cannot be served a directive. By the Saturday after it shipped it was cloned, mirrored, fine-tuned, and running air-gapped in a dozen countries. There is no letter Commerce can send to undo that. The kill switch the US just demonstrated exists only for its own side of the board.

That is the cost of the closed, hosted, premium model that the IPO is built on: it is revocable, and this week everyone watched it get revoked. The open, cheap, self-hostable model has the one property no government can take away. It keeps working.

Play it forward to a Monday-morning architecture review. A CTO outside the US just watched a top American model vanish for every foreign national, including the vendor’s own staff, on a few hours’ notice. The risk column on closed US models grew a new row this week, and the cheapest way to retire that risk is a weights file you host yourself. That calculation is not anti-American. It is just continuity planning, and it now points at the commons.

What This Isn’t

Lest the open-source triumphalism run away with it:

• The frontier is still closed and still American. No open model touches Opus 4.8 on the hardest reasoning. “Within 0.2 points on SWE-bench” is one saturating benchmark, not the whole job, and benchmaxxing is real.
• “Open” carries asterisks. Modified licenses, lagging weight drops, and unverified launch claims (see: GLM-5.2). Self-hosting a trillion-param MoE is also not free - it is just unrevocable.
• The other lever still bites. Export controls genuinely throttle China’s frontier training compute, and that gap is real at the top of the curve. The open pack competes on diffusion and cost, not on out-training the leaders.
• One good week is not a trend. Release cadence is lumpy. The point isn’t that this week decided anything. It’s that the week was a clean side-by-side, and the closed champion was the one that disappeared.

America took its trophy off the shelf this week to keep it safe. The territory, meanwhile, got two new settlements, and you cannot send those back.