Why is China open-sourcing its AI models?

China's best models are open; America's best are closed

On Artificial Analysis's Intelligence Index (version 4.3.2, checked 11 October 2026), the strongest models are closed and American, and the strongest open models are Chinese.

Model	Lab	Weights	Intelligence Index
Claude Opus 5.5	Anthropic (US)	Closed	58
GPT-6 Astra	OpenAI (US)	Closed	53
GPT-6.1 Sol	OpenAI (US)	Closed	52
MiMo-V2.6-Pro	Xiaomi (China)	Open	46
GLM-5.3	Z.ai (China)	Open	45
Kimi K3	Moonshot AI (China)	Open	44
DeepSeek V4.1 Flash	DeepSeek (China)	Open	39
GPT-6 Luna	OpenAI (US)	Closed	38
Gemma 4 31B	Google (US)	Open	15
gpt-oss-120b	OpenAI (US)	Open	12

The best open model sits 12 points below the best closed one. Google's and OpenAI's best open models score 15 and 12, and the highest-scoring US open model on the index, Thinking Machines' Inkling Small, scores 26. All are a long way under the Chinese open models. How fast the Chinese labs closed the rest of the gap is covered in has China caught up in AI.

The published explanations

Analysts give four main reasons, and they overlap.

State policy

A March 2026 working paper for the US-China Economic and Security Review Commission (Ngor Luong, "Two Loops") traces open source through Beijing's 2017 AI plan, the 14th Five-Year Plan in 2021, the State Council's "AI+" initiative in August 2025 and the draft 15th Five-Year Plan. Local governments in Beijing, Guangdong and Hangzhou have put millions into open-source AI. Premier Li Qiang said in 2025 that "China's innovation is open and open-source."

Influence and standards

A free model that developers adopt becomes the default they build on. Alibaba's Qwen had more than 100,000 derivative models on Hugging Face by March 2026 (the same USCC paper). Singapore's national AI program chose Qwen, and Malaysia said its sovereign AI system would run on DeepSeek (MIT Technology Review, 21 April 2026). This is the soft-power argument: countries that want an alternative to American platforms end up on Chinese ones.

Working around chip limits

With restricted access to the best US chips, Chinese labs get faster feedback by letting the world use and improve their models. MIT Technology Review called open-sourcing "a shrewd workaround." The USCC paper says the open ecosystem lets China "innovate close to the frontier despite significant compute constraints."

Commoditizing the complement

Joel Spolsky described the rule in 2002: "Smart companies try to commoditize their products' complements." Alibaba and Baidu sell cloud computing, so a free model that runs on their cloud brings them paying customers. On 11 October 2026 both companies' clouds were selling access to DeepSeek V4.1 Flash on OpenRouter, a model built by a rival lab.

Open releases also let labs build on each other. Moonshot's Kimi K2 (July 2025) runs on DeepSeek V3's architecture; its configuration file on Hugging Face names "DeepseekV3ForCausalLM."

My view: the government pushes its labs to open-source

I believe the Chinese government pressures its subsidized labs to open-source their weights. The published record supports part of that: national plans promote open source, and local governments fund it. Direct pressure on individual labs to release particular models is my opinion.

The effect, as I see it, is that anyone can run the models with no reliance on the company that made them. Competing labs also build on each other's released techniques, so the whole field improves fast.

What open weights means for you

When a lab publishes its weights, any company with the hardware can sell access to that model. On OpenRouter, a marketplace that resells many hosts, 29 companies were serving DeepSeek V4.1 Flash on 11 October 2026. They included DeepSeek itself, Alibaba, Baidu and US clouds such as Together, Fireworks and CoreWeave. OpenAI's open gpt-oss-120b had 21 hosts, and the most expensive charged about six times as much per output token as the cheapest. GPT-6 Luna, which is closed, had three (OpenAI, Microsoft Azure and Amazon Bedrock), all within 10% of OpenAI's own price.

For an ordinary user that competition shows up in two places.

First, price. Hosts undercut each other on the same model, so an app built on open models pays less to answer your question, and the market sets the price. Why that makes Chinese models so much cheaper is in why Chinese AI models cost less.

Second, data policy. Each host sets its own rules on keeping prompts and training on them, so an app can pick hosts whose rules suit it. Plainly does this: its account excludes providers that train on prompts, and your chats are never used to train AI. Some hosts still keep prompts for a period. The details are in does Plainly train on your chats.

Open weights also mean the model can't be taken away. On 10 September 2026 DeepSeek announced it would phase out its own V4-Pro API from 14 September, then reversed the decision "in response to user demand" and kept it running (DeepSeek's API change log). Had it gone ahead, the model would still be available: the V4-Pro weights are on Hugging Face, and on 11 October 15 other hosts on OpenRouter, including Microsoft Azure and Cloudflare, were serving it. The same freedom lets an app change models when a better one appears. Plainly runs one capable open-source model and switches when a better one comes out. You never pick.

Open weights vs open source: what's the difference?

Open weights means the trained model file is published under a license that lets you download and run it. Open source, under the Open Source Initiative's definition (version 1.0, October 2024), also requires the complete training code and detailed information about the training data. Most "open-source" models from China or the US publish only the weights, so open weights is the accurate term.

The license decides what you may do with the weights. Here are the most-downloaded open language models from each lab, with licenses from their model cards and Hugging Face's 30-day download counts as of 11 October 2026.

Model	Lab	On Hugging Face since	License	Downloads, last 30 days
Qwen3-0.6B	Alibaba	Apr 2025	Apache 2.0	30.8M
Gemma 4 26B A4B	Google	Mar 2026	Apache 2.0	11.9M
Llama 3.2 1B Instruct	Meta	Sep 2024	Llama 3.2 Community License	7.0M
GLM-5.3-Flash	Z.ai	Aug 2026	MIT	6.5M
gpt-oss-20b	OpenAI	Aug 2025	Apache 2.0	6.0M
DeepSeek V3.2	DeepSeek	Dec 2025	MIT	4.6M
Mistral 7B Instruct v0.3	Mistral AI (France)	May 2024	Apache 2.0	2.1M
Kimi K3	Moonshot AI	Jul 2026	Kimi K3 license (custom)	1.2M

Single-model counts favor small models that people run on their own machines. The most downloaded, Alibaba's Qwen3-0.6B, has 0.6 billion parameters, and the US labs make plenty of small models too. Across all labs, an MIT and Hugging Face study found Chinese open models took 17.1% of downloads in the year to August 2025, narrowly ahead of US models at 15.86%, the first time China led (MIT Technology Review, April 2026).

DeepSeek has published open weights since late 2023, when its first coding and chat models went up on Hugging Face. R1, in January 2025, is when the rest of the world noticed.

Doesn't the US release open models too?

Yes, below the frontier. Meta made Llama the best-known open model family from 2023, and released Llama 4 Scout and Maverick in April 2025 while holding back its largest model, Behemoth. Its next flagship, Muse Spark (codenamed Avocado), launched with closed weights in April 2026 (Bloomberg). Google publishes Gemma, now under Apache 2.0. Nvidia released its open Nemotron 3 Super model in March 2026 and the larger Nemotron 3 Ultra in June. In France, Mistral AI publishes many of its models under Apache 2.0.

The US government also backs open models. America's AI Action Plan, released in July 2025, tells the Commerce Department to "convene stakeholders to help drive adoption of open-source and open-weight models by small and medium-sized businesses." So the accurate split is this: the US frontier labs keep their best models closed and release smaller ones openly.

GPT-OSS vs DeepSeek

GPT-OSS is OpenAI's open-weights release, its first since GPT-2. Here it is against DeepSeek's current open model.

	GPT-OSS (OpenAI)	DeepSeek V4.1 Flash
Released	5 Aug 2025	10 Sep 2026
License	Apache 2.0	MIT
Size	120b: 117B total, 5.1B active. 20b: 21B total, 3.6B active	552B total, 8B to 16B active
Runs on	One 80GB GPU (120b), or 16GB of memory (20b)	Data-center hardware
Intelligence Index (v4.3.2)	12 (120b, high reasoning)	39 (max reasoning)
Downloads, last 30 days	20b: 5.95M. 120b: 3.85M	1.36M
Hosts on OpenRouter	21	29

Sources: Hugging Face model cards and API, Artificial Analysis leaderboard, OpenRouter's public model API, all checked 11 October 2026.

DeepSeek's model scores more than three times higher, and it is 13 months newer. GPT-OSS wins on downloads, probably because the 20b version runs on an ordinary computer. OpenAI itself positioned gpt-oss-120b at about the level of its o4-mini, well below its closed models.

I have used GPT-OSS a couple of times. It's okay, but it's far off OpenAI's own closed models, and I can't recommend it. The DeepSeek-class open models are the useful ones.

I believe GPT-OSS was a legal play. Elon Musk sued OpenAI in February 2024 over its move away from its nonprofit mission. OpenAI began as a nonprofit, added a capped-profit company in 2019 and became a public benefit corporation in October 2025. GPT-OSS arrived on 5 August 2025, while the case was running. No published source ties the release to the lawsuit, so treat that as my opinion. The published explanation is different: Sam Altman said in early 2025 that OpenAI had been "on the wrong side of history" on open models (cited in the USCC paper).

The trial ended on 18 May 2026, when a jury found that Musk had filed after the statute of limitations ran out and ruled for OpenAI (The Wall Street Journal). The question of whether OpenAI broke its mission was never decided. Musk has said he will appeal (Al Jazeera, 19 May 2026).

Where closed models still win

On the hardest problems, the closed US models lead by a clear margin: Claude Opus 5.5 scores 58 on the index against 46 for the best open model. For difficult code and research-grade questions, that gap is worth paying for. For everyday questions, writing and planning, the open models are close enough that price and choice of host matter more. Open Chinese models also follow China's content rules on political topics (MIT Technology Review, April 2026), which is covered with the privacy questions in are Chinese AI models safe.