A few days ago, while researching text-to-video models, I discovered something that differed significantly from my original understanding: in the directions of text-to-image and text-to-video, Chinese companies' presence in the open source community is far weaker than in language models.
I had always assumed Chinese models dominated across all fronts. But upon closer inspection, that's not the case. This led me to review the changes in the open source community over the past three years.
Model Open Source Is Not Software Open Source
First, let's discuss something many people may not have considered.
With open source software, once the source code is released, there are no secrets left—you can rebuild an identical copy from scratch. Models are different. When models are open sourced, what you typically get are the weights and inference scripts. The core elements—training methods, training data, engineering details—usually remain undisclosed.
What can you do with the weights? Deploy them for inference, fine-tune them. Want to reproduce from scratch? Nearly impossible. So "open source" for models and "open source" for software were never the same thing from the start.
This leads to a question that was debated extensively in the early days: since model open source differs from software open source, what is the point of open sourcing models?
The greatest benefit of open source software is community collaboration—developers worldwide fixing bugs and adding features together. But after open sourcing a model, very few individuals or institutions actually have the capability to participate in model development. You need massive compute, data, and training infrastructure. The vast majority of people can only run inference with the weights.
At the time, Robin Li said open source large models made no sense, and for a while I thought he had a point. If the primary driver of community collaboration doesn't apply, what's the point of open sourcing?
Subsequent events answered this question.
The Era of LLaMA
When ChatGPT emerged at the end of 2022, large models entered the public consciousness. Before this, the open source world was rather monotonous. OpenAI's GPT-2 was fully open sourced in 2019, but by GPT-3 it had shifted to an application-based API access model. I remember wanting to use GPT-3 at a hackathon and discovering I had to email to request API access. Essentially, it was already closed source.
In 2023, the open source community was basically dominated by Meta's LLaMA. LLaMA 1 came in February 2023, LLaMA 2 in July. Every time LLaMA released a new version, a batch of Chinese models would follow up with upgrade announcements—this was the so-called "Hundred Model War" (Bai Mo Da Zhan), with the pace dictated by LLaMA.
Chinese models open sourced at this stage were, frankly speaking, for marketing purposes. The models released were relatively small. Zhipu's GLM-6B was the earliest representative; many people's first exposure to private deployment of large models started with it. I remember a friend choosing a model at the time, and I wondered why he picked a Chinese model. He said it was made by Tsinghua and had a good reputation. Baichuan open sourced a 14B model. 01.AI (Zero One Everything), founded by Kai-Fu Lee, open sourced Yi-34B in November 2023, which was considered relatively large among Chinese open source models at the time. Shanghai AI Laboratory also continued working on the InternLM (Shusheng) series.
Everyone's strategy was the same: open source small models for promotion, keep large models closed for commercialization.
The Entry of Qwen
In 2024, this balance was broken by Qwen.
Starting mid-2024, Alibaba's Qwen made intensive efforts, releasing models from several billion to 72B parameters, all performing well and fully open sourced. Previously, everyone assumed large models wouldn't be open sourced; suddenly someone released a highly effective large model openly.
Although LLaMA had significant international influence, its Chinese capabilities were always lacking, requiring secondary training for practical use. Qwen could be used almost out-of-the-box for Chinese scenarios, quickly replacing LLaMA's position in the Chinese open source community.
By the end of 2024, the Qwen series had become the de facto standard for Chinese open source models. Closed source models felt pressure from open source for the first time.
DeepSeek Flipped the Table
Qwen did the best within the existing rules. DeepSeek changed the rules entirely.
DeepSeek entered the scene in 2024 with a simple approach: open source upon release, with extremely thorough technical reports. At the end of 2024, V3 was released—hundreds of billions of parameters, excellent performance, open sourced immediately upon release. At that time, few had seen models of this scale released openly.
But what really caused an explosion was R1 in January 2025.
OpenAI had just launched the O1 reasoning model in September 2024, and DeepSeek's R1 came out before the Spring Festival (Chinese New Year). Its reasoning capabilities were very close to the top closed source models at the time—not quite on par, but the gap was surprisingly small. And it was fully open sourced on day one.
Previously, everyone maintained an order of "small open source, large closed source." What DeepSeek open sourced was better than many companies' best closed source models, and this order collapsed immediately.
LLaMA 4 is another footnote. Meta spent a long time training a massive model to regain its position in the open source community, releasing it in April 2025, but it flopped. Performance fell far short of expectations, and cheating on benchmarks was exposed. Later, Yann LeCun himself admitted that "results were fudged," and Zuckerberg lost confidence in the entire GenAI team. LLaMA 4 is basically unused, and the LLaMA series' position in the open source community ended there.
Day-0 Open Source Became the Industry Standard
After DeepSeek, Chinese model companies turned to day-0 open source one after another—their best models open sourced on the day of release.
Kimi open sourced K2 in July 2025, a trillion-parameter MoE model. MiniMax open sourced M2.5. Zhipu continued iterating the GLM series. Wave after wave, the quality of open source models kept rising.
Today, if you look at the international open source community for language models, the leaderboards are almost entirely Chinese models. Qwen, DeepSeek, GLM, MiniMax, Kimi—the presence of overseas models has become very weak.
What DeepSeek did wasn't just contribute a model. It changed how the entire industry plays its hand.
But the Winds Are Shifting Again
However, the fervor for this wave of day-0 open source is cooling down.
DeepSeek's last open source release was V3.2 in December 2025—more than four months ago. V4 has been rumored for a long time but hasn't appeared. During this window, everyone's strategies have started to loosen.
Qwen 3.6 Plus was released at the end of March 2026, not open sourced, API-only. This is the first time a flagship Qwen model wasn't open sourced. Zhipu's GLM-5.1 was also released closed source first, although they just announced they would open source the weights in the past couple of days. Many companies' latest multimodal models are no longer open sourced either.
It seems we've returned to that original question: what's the point of open source? When competitive pressure decreases, the answer may change again.
Text-to-Image and Text-to-Video Are Still Waiting
Returning to that initial surprising discovery.
The text-to-image open source community is still dominated by overseas models. The most widely used are Stability AI's Stable Diffusion series and Black Forest Labs' FLUX series. Chinese models have made some progress—Qwen released Qwen-Image, Tencent has Hunyuan Image 3.0, Zhipu has GLM-Image. But compared to the situation with language models, the difference is vast.
Text-to-video is the same. Alibaba's last open source text-to-video model was Wan 2.2 in July 2025—almost 9 months ago with nothing new since. The recently popular open source text-to-video model is LTX-2, from Israeli company Lightricks, which open sourced weights in January 2026.
This is a completely different world from language models. On the language model side, Chinese models have filled the entire open source community; on the text-to-image and text-to-video side, it still looks more like 2023: overseas models dominate, Chinese models appear sporadically.
What Are We Waiting For
Everyone is waiting for DeepSeek V4.
But we're waiting for more than just a model. DeepSeek previously proved something: a sufficiently strong model fully open sourced on the day of release can change the strategic direction of an entire industry. This happened with language models; it hasn't happened yet with text-to-image and text-to-video.
I sometimes half-jokingly think that maybe DeepSeek just needs to make another move. But then again, DeepSeek itself hasn't released a new model in over four months either.