How China Is Driving the Open-Weight AI Race
DeepSeek, Z.ai, Moonshot and Qwen are not just releasing bigger open-weight models. Together they are changing the pace, economics and developer ecosystem of open AI.
For most of the generative-AI era, the global race has centred on OpenAI, Anthropic, Google and Meta, with frontier intelligence delivered mainly through proprietary services and APIs. Underneath it, a second contest has accelerated: becoming the foundation of open-weight AI, where trained weights can be downloaded, adapted and deployed outside the company that created them. In 2026, that race has developed a distinctly Chinese centre of gravity.
Cloud Codes grouped seven releases or announcements into one remarkable burst: Meta Muse Glimmer 30B, NVIDIA Nemotron 3 Ultra, Alibaba's Qwen3.8-2.4T-A95B, DeepSeek V4, Qwen3.8-27B, Z.ai's GLM-5.3 and RedNote's 280B model. Five came from Chinese labs; Meta and NVIDIA were the Western open-weight counterpoint.
That list is useful as a snapshot, not as a literal release ledger. DeepSeek's April V4 was a preview, with DeepSeek-V4-Flash-0731 becoming the official Flash release on 31 July. Z.ai announced GLM-5.3 on 14 August but said public weights would follow after further safety work. The more durable story is the pace: several Chinese labs are now taking turns at the frontier, while Alibaba is releasing at both ends of the market.
China is not producing a single open-weight challenger. It is producing a tournament.
The open frontier is moving east
Hugging Face's State of Open Models: Summer 2026 report makes the shift unusually visible. In almost every month it examined, the largest high-performing open model from a Chinese lab was larger than the largest US counterpart. China's monthly ceiling ranged from roughly 754 billion to 2.78 trillion parameters.
Size is not intelligence. Many of these models use designs that activate only part of the network at a time. The useful signal is simpler: Chinese labs are repeatedly willing to release very large models as downloadable weights.
DeepSeek's April preview introduced V4-Pro at 1.6T total parameters with 49B active and V4-Flash at roughly 284B total with 13B active. On 31 July, V4-Flash-0731 became the official Flash release, superseding the preview while keeping the weights under MIT. (DeepSeek V4-Flash-0731)
Z.ai's GLM-5.2 is a 753B MIT-licensed model with a one-million-token context window, aimed at long-running engineering work. (GLM-5.2)
Moonshot AI's Kimi K3 goes further at 2.8T total parameters, 104B active, with image input and a one-million-token context window. Alibaba's hosted Qwen3.8-Max is 2.4T/95B-active, followed by the downloadable Qwen3.8-2.4T-A95B checkpoint. The open checkpoint is text-only and thinking-only, so it is not identical to the hosted multimodal product. (Kimi K3) (Qwen3.8)
> CLOSED FRONTIER / OPEN-WEIGHT STACK
Qwen's stronger argument is not size
The most consequential number in this story may not contain the letter T. It is 151,448: the number of Qwen-derived repositories Hugging Face counted in its Summer 2026 report, 2.6 times Meta's total derivative footprint and 4.7 times Llama specifically. Through the first seven months of the year, roughly 180 to 210 new Qwen derivatives appeared each day. That shows what developers chose to fine-tune, quantise, package and build on.
Hugging Face also counted roughly 2.045 billion Qwen downloads in 2026 across repositories with declared parameter counts. More revealingly, only 3% of 2026 Hub download volume went to models above 70B parameters. The giant models get the headlines; smaller models get most of the traffic.
That is why Qwen3.8-27B may be the most important open-weight model of 2026. Not because it is the largest, but because it sits where deployment actually happens: a dense 27B model, Apache 2.0 licensed, natively multimodal and far more realistic to run, test and adapt than a trillion-parameter flagship. (Qwen3.8-27B)
The community multiplies that advantage. Hugging Face counted 28,531 Qwen GGUF conversions, only 54 published by Qwen itself. Developers are doing most of the packaging work. That is how a model family becomes infrastructure rather than a launch announcement.
Open weights change where value accumulates
Why spend enormous sums training a model and then make the weights downloadable? Because the weights do not have to be the only thing that makes money. Value can move into APIs, cloud services, support, hardware and distribution.
Open weights are not risk-free. Licensing can tighten at the top: Kimi K3 and Qwen3.8's 2.4T checkpoint add conditions for some very large commercial users and Model-as-a-Service businesses. Qwen3.8-27B, by contrast, is Apache 2.0.
There is also continuity risk. Qwen technical lead Junyang Lin stepped down in March without publicly giving a reason. The Financial Times later reported disagreement inside Alibaba over the cost of open-model development and a stronger push toward cloud revenue. Qwen3.7-Max and Qwen3.7-Plus appeared as hosted models, but no comparable Qwen3.7 open-weight family followed. Qwen3.8 later returned to downloadable weights. Once you download a model, that copy is yours; the upgrade path still belongs to the organisation training the next generation.
The harder governance trade-off is trust. A closed Western model asks an organisation to trust the provider, its contracts and the jurisdiction around it; access can change through policy, legal requirements, sanctions or product retirement. A Chinese open-weight model gives much more control over where it runs, but its training data and provenance may be too opaque for some regulated or risk-sensitive organisations. The problem is not unique to Chinese models.
Closed models offer provider accountability but less operational control. Open weights offer more operational control but do not automatically solve provenance. Neither removes trust; they move where the trust sits.
The community is making giant models usable
Quantisation and tools such as llama.cpp reduce memory use and widen the hardware that can run open models. They do not turn Kimi K3 into a laptop model, but they lower the barrier to experimentation.
Qwen again stands out. Hugging Face recorded roughly 39.6 million Qwen GGUF downloads per month, compared with 20.8 million for Gemma and 7.5 million for Llama. Llama-derived GGUF repositories slightly outnumbered Qwen's, so the difference was not simply that more Qwen conversions existed.
Labs release weights; the community adapts them; inference tools and hardware vendors add support; developers build applications. The competitive unit is increasingly the ecosystem, not the checkpoint.
From models to agent harnesses
The 2026 releases are designed for sustained work as well as isolated answers. DeepSeek, GLM, Kimi and Qwen all emphasise coding, tool use or long-running tasks.
Earlier chatbots mostly answered a prompt and stopped. An agent is given an outcome, works through a sequence of steps, uses tools, checks the result and adjusts when something goes wrong. That makes the software around the model much more important.
DeepSeek publishes DeepSeek Harness under MIT, with interchangeable models, tools, sandboxes and agent loops. A separate June paper, HarnessX, reported an average 14.5 percentage-point improvement across five agent benchmarks, with gains up to 44 points in some tested setups. That is not a universal 44-point boost; it is evidence that the surrounding software can materially change what an agent can do. (DeepSeek Harness) (HarnessX)
The deployable unit is increasingly model + harness + inference stack. Open weights let an organisation control more of that stack instead of depending entirely on one API provider.
China has not "won AI"
None of this supports the sweeping claim that China has won artificial intelligence. Parameter count is not a quality score. Vendor benchmarks need independent scrutiny. Downloads and derivative counts are ecosystem signals, not market share.
The Western open ecosystem is active too. NVIDIA's Nemotron 3 Ultra is 550B total / 55B active; Arcee's Trinity-Large is roughly 398B; Meta's Muse Glimmer 30B is Apache 2.0; and Poolside's Laguna S 2.1 is 117.6B / 8.5B active. Thinking Machines' Inkling is 975B / 41B active, Apache 2.0 and trained from scratch. Thinking Machines also says its first supervised fine-tuning stage used synthetic data from open models including Kimi K2.5. Chinese open models are influencing parts of the Western open stack as well. (Laguna S 2.1) (Inkling)
Chinese labs are setting much of the tempo at the high end of open weights, while Qwen is turning that tempo into developer habit. The model that best expresses the strategy may be Qwen3.8-27B: practical enough to run, permissively licensed, surrounded by a huge derivative ecosystem and positioned where most open-model usage actually happens.
For years, the industry assumed the strategic prize would belong to whoever built the world's most capable individual model. The open-weight race suggests another possibility: the winner may be the model family, harness and distribution stack that everybody else decides to build on.
In 2026, that contest looks very different from the proprietary AI race most of the world has been watching.
Sources
Factual sources checked 18 August 2026.
- Hugging Face, State of Open Models: Summer 2026 Observations, 14 Aug 2026
- DeepSeek, DeepSeek V4-Flash-0731, 31 Jul 2026
- DeepSeek, DeepSeek Harness
- Z.ai, GLM-5.2
- Z.ai, GLM-5.3 announcement
- Moonshot AI, Kimi K3
- Moonshot AI, Kimi K3 licence
- Qwen, Qwen3.8-2.4T-A95B
- Qwen, Qwen3.8 licence
- Qwen, Qwen3.8-27B
- Alibaba Cloud, Qwen3.7-Max
- Alibaba Cloud, Qwen3.7-Plus
- Financial Times, Alibaba shifts towards revenue over open-source AI, Apr 2026
- Meta, Muse Glimmer 30B
- Poolside, Laguna S 2.1
- Thinking Machines Lab, Inkling
- Thinking Machines Lab, Inkling model card
- NVIDIA, Nemotron 3 Ultra
- Darwin Agent Team, HarnessX, Jun 2026
- Cloud Codes, The Week Open Source Won: 7 Open Models in 7 Days, used as a research lead; individual claims checked against direct sources
Something wrong here? Corrections are welcome at hello@devfactory.mu and will be noted on the article.