Four Models, One Week, Zero Licensing Fees: China Just Open-Sourced the AI Race

Published February 16, 2026

The US spent the last two years trying to slow down Chinese AI development with chip export controls. Here’s how that’s going:

Alibaba RynnBrain (Feb 10) — Robotics AI. Beat Google and NVIDIA across 16 benchmarks. Open source. Free.
Zhipu AI GLM-5 (Feb 11) — 745B parameter LLM trained entirely on Huawei chips. Not a single NVIDIA GPU touched this thing. Open source, MIT license.
MiniMax M2.5 (Feb 12) — Matches Claude Opus 4.6 on coding benchmarks at one-hundredth the price. Open weights.
ByteDance Seedance 2.0 (Feb 12) — Video generation so good that Disney sent a cease-and-desist within 24 hours.

That’s one week. Four models. Three of them are completely free to use, modify, and deploy however you want.

Meanwhile, the American releases this month: Claude Sonnet 5 at $3 per million tokens and GPT-5.3 Codex behind enterprise pricing. One side is building in public. The other side is building paywalls.

The Numbers That Matter

Forget benchmarks for a second. Look at the economics.

MiniMax M2.5 runs at 100 tokens per second — twice as fast as other frontier models. API pricing: $0.15 per million input tokens. Claude Opus 4.6 charges $15 per million input tokens. That is a 100x price difference for comparable coding performance.

MiniMax claims 80% of their internal code is now written by M2.5. If that number is even half true, they have built a self-sustaining development flywheel at a fraction of what Western labs spend.

GLM-5 is the one that should keep Washington up at night. It was trained without a single NVIDIA GPU — entirely on Huawei Ascend chips. It outperforms Claude Opus 4.5 on Humanity’s Last Exam. It has the lowest hallucination rate in the industry, achieved through a novel reinforcement learning technique they call “slime” (their word, not mine).

US chip export controls were supposed to slow Chinese AI development by years. GLM-5 suggests the controls created a constraint that forced innovation. The restriction became the strategy.

RynnBrain is playing a longer game. Robotics AI — “physical AI” — is where the real-world applications live. Not chatbots, not content generation, but robots that can navigate a kitchen, predict object trajectories, and execute multi-step tasks in messy physical environments. Alibaba’s model beats Google’s Gemini Robotics and NVIDIA’s Cosmos Reason, and they released it for free on Hugging Face. Any robotics startup in the world can now build on it.

This Is Not an Accident

Chinese open-source models now account for 30% of global AI downloads. Two years ago it was less than 10%. US models? 15.7%. Alibaba’s Qwen family has overtaken Meta’s Llama in cumulative Hugging Face downloads.

You’ve seen this playbook before. It’s the same one that made Android dominant, Linux ubiquitous, and Chinese manufacturing inescapable: give it away for free until the expensive option looks like a scam. It works every time because it turns out most people prefer free things that work over expensive things that also work.

Lian Jye Su, Chief Analyst at Omdia, put it plainly: “DeepSeek showed the industry that you can create a very good model even when you’re resource-constrained. The combination of open-source access, strong reasoning capabilities and low deployment costs has become a defining model for how Chinese vendors now approach foundation models.”

A year after DeepSeek R1 flipped the “closed models will win” thesis, the February 2026 wave proves it was not a fluke. It is a sustained national strategy. Even Baidu’s Robin Li — who was publicly anti-open-source — reversed his position and started opening models.

What This Means for You

If you are a developer, this is unambiguously good news. More models, more competition, lower prices. MiniMax M2.5 at $0.15/M tokens means you can build things that were economically impossible six months ago.

If you are a startup building on Claude or GPT APIs, you should be testing these models today. Not because they are necessarily better, but because your unit economics just got disrupted. When a comparable model costs 5% of what you are currently paying, “we use the best model” is no longer a sufficient justification for your burn rate.

If you are Anthropic or OpenAI, the competitive moat is shrinking by the week. The response cannot be “our model is slightly better on benchmarks.” It has to be ecosystem, tooling, reliability, and trust — the things that are hard to replicate even when the raw model capability is commoditized.

If you are a US policymaker who thought chip export controls would maintain a durable AI advantage: GLM-5 was trained on domestic Chinese hardware and competes with frontier Western models. The controls bought time. They did not buy victory.

The Uncomfortable Question

There is a real concern here, and it is not about capability. It is about governance.

Chinese open-source models come with the same censorship constraints as their closed counterparts. Ask GLM-5 about Tiananmen Square and you will hit a wall. The models are open in code but closed in values.

This creates a strange dynamic: the most accessible AI in the world may also be the most ideologically constrained. Developers can fine-tune away the censorship for their own deployments, but the default behavior shapes the billions of interactions that happen through the base models.

Open source is not the same as open governance. A model can be free to use and still serve a particular vision of what can be said.

Where This Goes

The AI race is not US vs. China. That framing is outdated.

It is closed vs. open, with geography as a proxy. American labs are mostly closed (with the exception of Meta). Chinese labs are mostly open. European labs are… well, European.

The February 2026 model wave crystallized something that has been building for a year: the open-source approach is not just competitive. It is winning the distribution game. When 30% of global AI downloads come from Chinese open-source models, that is not just market share. That is influence.

The Western response needs to be more than bigger models behind higher paywalls. It needs to be a credible open alternative — one that is both free AND reflects democratic values.

Meta is trying. Everyone else is watching.

Kyber Intel covers AI from the individual’s perspective. Not the corporation’s. Not the government’s. Yours. Follow along at kyberintel.com.