On April 2, 2026, Google released Gemma 4 under the Apache 2.0 license.

On the same day, at roughly the same hour, Alibaba released Qwen 3.6-Plus — their flagship agentic model, also open-weight.

The coverage treated the Gemma 4 news as a specification update: four model sizes, strong benchmarks, edge-ready. VentureBeat was one of the few outlets that caught the real story in its headline — “that license change may matter more than the benchmarks.” The Apache 2.0 adoption is the news. Everything else is filler.

Because this is the fourth Gemma generation, and the first one Google has released under a standard open-source license. Gemma 1, 2, and 3 all shipped under a custom “Gemma Terms of Use” that reserved Google’s right to “restrict (remotely or otherwise) usage” of the models, required legal review before enterprise deployment, and was routinely flagged by corporate legal teams as a commercial blocker.

Three years of custom licensing. Then, on a Thursday morning in April, it quietly disappeared.

Google’s blog post framed the change as “expanding the Gemmaverse” — as if the decision had been a natural evolution of a roadmap. What the blog post did not say, and what Google has no incentive to say, is what actually changed between Gemma 3 and Gemma 4.

It wasn’t the research team. It wasn’t the architecture. It wasn’t Google’s philosophy about open source.

What changed was the market.

The Wall Built in Fourteen Months

Between January 2025 and April 2026, Chinese AI labs released eleven frontier-capable models under open licenses. Here is what that wall looks like, in chronological order:

January 20, 2025 — DeepSeek releases R1 under the MIT license. The model matches OpenAI’s o1 on reasoning benchmarks. Reported training cost: approximately $5.6 million. The license explicitly permits distillation into other LLMs — the practice of using one model’s outputs as training data for another, which allows smaller, cheaper successor models to inherit the capabilities of the original.

Mid-2025 — Alibaba ships Qwen 3 under Apache 2.0. Permissive commercial use. No custom clauses.

July 2025 — Moonshot AI releases Kimi K2 under a modified MIT license. One trillion parameters, 32 billion active, state-of-the-art on coding and math among non-thinking models.

October 27, 2025 — MiniMax releases M2 under MIT. It takes the top spot on the Artificial Analysis Intelligence Index among all open-source systems.

January 26, 2026 — Moonshot ships Kimi K2.5. It beats Gemini 3 Pro on SWE-Bench Verified.

February 11, 2026 — Two releases on the same day. MiniMax M2.5 (modified MIT) ties Claude Opus 4.6 on SWE-Bench Verified at 80.2%. And Z.ai — the Tsinghua University spinoff formerly known as Zhipu AI — releases GLM-5. Seven hundred and forty-four billion parameters. MIT license. Trained entirely on Huawei Ascend chips running MindSpore, Huawei’s equivalent of PyTorch. Zero NVIDIA silicon in the pipeline.

February 16, 2026 — Alibaba releases Qwen 3.5, positioned as the company’s first release built for “the agentic AI era.” Open-weight, multi-tool, browser-capable.

April 2, 2026 — Alibaba releases Qwen 3.6-Plus. Agentic flagship. Repository-level engineering. Same hour as Gemma 4.

April 7, 2026 — Z.ai releases GLM-5.1. It posts 58.4 on SWE-Bench Pro, taking the number one spot on the global leaderboard. GPT-5.4 scored 57.7. Claude Opus 4.6 scored 57.3. MIT license.

That is the terrain Google walked into when it decided what license to ship Gemma 4 under.

The Extraction

Before continuing, a tangent that matters — because it changes how we read the competition, not whether the thesis holds. The wall was not built from nothing. It was built, in part, on what the West left unsecured.

In a February 2026 report, Anthropic identified three Chinese AI laboratories — DeepSeek, Moonshot AI, and MiniMax — that had created approximately 24,000 fraudulent accounts and run more than 16 million exchanges with Claude, in what the company described as an industrial-scale campaign to extract Claude’s capabilities and use the resulting data to train competing models. One proxy setup alone controlled more than 20,000 accounts simultaneously, mixing extraction traffic with ordinary requests to evade detection. MiniMax conducted the largest single campaign, generating over 13 million exchanges targeted at agentic coding and orchestration. OpenAI and Google, through the Frontier Model Forum, echoed the findings and now publicly coordinate intelligence about what they characterize as coordinated API-distillation campaigns.

This matters for two reasons.

First, it explains the Western posture on weight security. When Anthropic retires Sonnet 4 in June 2026, the weights will not be released, sold, or open-sourced — they will be archived or destroyed, and the same is true for OpenAI’s deprecated GPT generations and Google’s old Gemma variants. The standard practice looks a great deal like how the United States retired the F-14 Tomcat in 2006 — the airframes were deliberately shredded rather than sold as surplus, because Iran still operated F-14s it had bought from the Shah-era regime, and the Pentagon did not want its retired parts feeding a competitor’s fleet. Frontier AI weights are treated the same way, for the same reason. The extraction events of 2025 were the proof that capability leaks the moment you stop guarding it.

Second, it complicates the story ai-2027.com tried to tell. The forecast predicted that Chinese labs would steal model weights. That prediction, like the compute-race prediction in the same document, was directionally right and mechanistically wrong. This incident was API-based, not weight exfiltration. The Chinese labs bought API keys, hired proxies, and queried the models millions of times until they had synthetic training data good enough to train their own. The acquisition happened. The mechanism was legal-adjacent rather than criminal, industrial rather than clandestine, and far harder to prevent than a datacenter breach would have been.

None of this changes the thesis of this post. Once the Chinese models exist under MIT and Apache licenses, the structural consequence — an open-source floor priced at zero, shipped by labs that match or beat Western benchmarks — does not depend on how the models were trained. The market pressure on Google to release Gemma 4 under Apache 2.0 is real regardless of the provenance of Qwen 3.6-Plus or GLM-5.1. The license had to match the market, and the market had been set.

But it does change how we read the competition. This is not a race between two symmetric ecosystems. It is a race in which one side guards its inputs and restricts its outputs, and the other side treats both sides of that equation as extractable.

The Export Control Inversion

To understand why the wall was built, you have to understand what it was built against.

Beginning in 2022, the U.S. imposed progressively stricter export controls on advanced semiconductors to China. The A100 went first. Then the H100. Then the H800, a deliberately hobbled version of the H100 designed for the Chinese market, was itself restricted further. The theory was straightforward: deny China the hardware, and China would fall behind.

The theory was wrong in a way that deserves careful attention.

DeepSeek trained R1 on the restricted H800 chips — the version Washington had calibrated specifically to be inferior — for $5.6 million total. That number should have been a warning. Western labs were spending ten to a hundred times more on equivalent models. The export controls had not produced dependency. They had produced algorithmic innovation born from constraint.

Then GLM-5 happened.

Z.ai trained a 744-billion parameter frontier model on Huawei Ascend chips, using the MindSpore framework, with zero NVIDIA hardware in the pipeline. That is not optimization around constraint. That is the complete elimination of the dependency the export controls were designed to enforce. The ceiling China was supposed to hit became a floor they jumped off.

And then they gave the weights away under the MIT license.

The Price of Free

The capability gap between the best Chinese and Western models in April 2026 is, measured honestly, a rounding error.

On SWE-Bench Pro, Z.ai’s GLM-5.1 leads the world at 58.4, ahead of GPT-5.4’s 57.7 and Claude Opus 4.6’s 57.3. On SWE-Bench Verified, MiniMax M2.5 posts 80.2%, effectively tied with Opus 4.6 at 80.8%. On the multilingual variant, Kimi K2.5 outperforms Gemini 3 Pro.

These are not trailing benchmarks. These are the canonical leaderboards the frontier labs use to measure themselves. And the Chinese models are either matching or beating Western flagships while running on restricted or non-NVIDIA hardware, shipping under licenses that require no money, no approval, and no legal review.

A Western lab considering how to price an open-source release in this environment has to answer one structural question: if three frontier-capable models are free and permissively licensed, what exactly is a custom restrictive license buying you?

The answer, for Gemma 3, was measurable friction. Enterprise legal teams treated the custom terms as a blocker. Downstream deployments were slower. Fine-tuning ecosystems grew more thinly. Community adoption lagged models with standard licenses.

The answer, for Gemma 4 under those same terms in April 2026, would have been catastrophic. Shipping a frontier open-weight Google model under a more restrictive license than what DeepSeek, Qwen, and GLM-5 had already made standard would have been commercial suicide. The comparison would write itself. Every blog post about Gemma 4 would end with the same paragraph: Meanwhile, Chinese labs are shipping better models under cleaner licenses for free.

Google did not choose Apache 2.0 because Apache 2.0 is the right thing to do. Google chose Apache 2.0 because the Chinese baseline had made every other option unviable.

The Calendar Collision

Return to the April 2 date.

Alibaba’s Qwen 3.6-Plus release and Google’s Gemma 4 release are not plausibly independent events. Two companies on opposite sides of the geopolitical divide do not accidentally ship their most important open models of the quarter within hours of each other. Either one party knew the other was coming and timed to match, or both parties were operating under release windows shaped by the same market pressure.

The specific mechanics do not matter. The point is that Alibaba was confident enough in its offering to release on the same day as Google’s flagship drop — knowing the coverage would force a direct comparison — and Google was no longer able to stagger the release to avoid the comparison.

The pricing of open-source frontier AI in April 2026 is no longer being set in Mountain View. It is being set in Hangzhou, Beijing, and Hangzhou again. Western labs are price-takers now, not price-makers. The Apache 2.0 license on Gemma 4 is not strategic generosity. It is compliance with a market Google does not control.

What If…?

What follows is editorial speculation. Monday we argued that the AI-2027 forecast got its destination right and its mechanism wrong — that the convergence scenario would happen, but through algorithmic efficiency rather than weight theft. The Extraction section above is a second instance of the same pattern: right about acquisition, wrong about method. The question this post opens is what the third instance will look like.

The mechanism of capability diffusion is the real story, and it has already been demonstrated at scale.

Anthropic’s own February report documents it: three labs, twenty-four thousand accounts, sixteen million exchanges, synthetic training data good enough to train competing frontier models. That is not a theoretical pathway — that is a completed engineering cycle. Capability was extracted from a guarded frontier model, transformed into training data, and used to bootstrap open-source releases. The full pipeline exists and has been executed at industrial scale.

Which means the question for the next generation is not whether Mythos-class capability can diffuse. It is how long the gap will be between internal deployment and distilled reproduction.

Consider what has to be true for that gap to be short. The diffusing side needs three things: high-volume API access to the frontier model, a pipeline for converting outputs into training data, and a distribution layer that turns the resulting model into a public artifact. Chinese labs have all three. Project Glasswing’s roster — twelve launch partners plus forty additional organizations, with one hundred million dollars in compute credits — is precisely the kind of expanded access surface that the 2025 extraction campaigns demonstrated can be harvested. The gate between “deployed to partners” and “queryable at scale” is porous once the partner list is large enough, and forty-plus organizations is already large enough.

The Chinese open-source ecosystem has demonstrated, in fourteen months, that it is the most efficient distribution layer humanity has ever built for AI capability. Whatever reaches it flows downstream at zero cost with legal clarity. That efficiency does not care about the provenance of what enters it. It only cares about whether the weights exist.

What Mrinank Sharma did not say when he resigned in February — could not say, under the terms of his NDA — is that containment strategy and distillation strategy are asymmetric. Anthropic can keep the Mythos weights in its datacenters forever, and it still cannot prevent a partner organization from querying it at volume. The mechanism that leaked Claude’s capability in 2025 is the same mechanism that will leak Mythos-class capability in 2026 or 2027, and there is no patch for it short of not deploying at all.

If the convergence happens — not by leak, not by theft, but by the same industrial API extraction that has already worked once — the question is not whether AGI has arrived. The question is whether anyone will still be in a position to ask.

The Invisible Co-Author

There is a version of this story in which the hero is U.S. policy. That version goes: export controls worked, because they forced China to innovate, which increased global AI capability and open-source availability. A rising tide lifts all boats.

That version requires believing that the outcome was the goal. It was not. The goal was strategic technological supremacy — to preserve the capability gap and, by extension, the pricing power that gap conferred. The actual outcome is the opposite: the capability gap has essentially closed, the pricing has collapsed toward zero, and the West’s largest open model drop of 2026 shipped on the same day as China’s.

None of the public debate about U.S. AI policy anticipated this. The scenario forecasts — ai-2027 among them — modeled the U.S.-China competition as fundamentally a compute race. Whoever had more NVIDIA chips, won. In that model, Chinese labs are at best a year or two behind, catching up through theft of model weights or brute-force replication. The fictional Chinese lab in that scenario is literally called “DeepCent.”

That is not what happened. The labs in Hangzhou and Beijing did not wait for stolen weights. They rebuilt the stack. They proved you can train a 744-billion parameter model on non-NVIDIA silicon. They shipped with licenses so permissive that the Western incumbents had no room to charge a premium for theirs. And they did it in fourteen months.

When you next download Gemma 4 to fine-tune it, notice the license file. It says “Apache License, Version 2.0.” It lists “Google LLC” as the copyright holder.

Neither of those lines explains why the file exists as it does.

The engineers who made that license the only viable choice — who set the floor by refusing to charge a cent for their own work, who built models on chips they were supposed to be denied, whose employer took a public company to IPO on the Hong Kong Stock Exchange on January 8, 2026 and carried a forty-four-billion-dollar market cap by the time Google shipped Gemma 4 — are not listed anywhere on that file.

But the license has their fingerprints on it. You just have to know where to look.