(Yicai) Aug. 13 -- Chinese artificial intelligence developers are increasingly making the weights of their flagship AI models available for public download, allowing users to deploy the models on their own infrastructure and potentially cut costs, in a move that challenges closed-source competitors.

DeepSeek launched the official version of its open-source DeepSeek V4 Pro model today while Alibaba Group Holding released the model weights for Qwen3.8-2.4T-A95B the same day, marking the first time the Hangzhou-company has made the weights of one of its flagship Max-tier models publicly available.

Open weights allows users to deploy, run inference, fine-tune or modify the model on their own hardware without relying on the developer’s Application Programming Interfaces. This differs from fully open-source models, which usually disclose training data, complete source code and the training process.

The latest wave of open-source or open-weight updates to models by leading AI firms is part of a broader trend rather than an isolated development. Just a few days ago, the US’ Meta released its open-weight Muse Glimmer model and said it plans to release the weights of its flagship Muse Spark 1.2 model. Over the past several weeks, Chinese AI models, including Kimi K3, Zhipu AI’s GLM-5.2 and MiniMax H3, have also undergone successive updates.

The latest generation of open models is placing a greater emphasis on agentic capabilities. In the AI coding-agent benchmark DeepSWE, the official version of DeepSeek V4 Pro achieved a score of 62.7, up from 12.8 for its preview version. Alibaba’s Qwen team has demonstrated that Qwen3.8-Max can autonomously code for around 16 days, independently reproduce and improve on academic research papers, and run a virtual e-commerce business through more than 2,000 interactive sessions.

Chinese-developed open models accounted for 41 percent of global downloads, surpassing US-developed models for the first time, according to UA AI platform Hugging Face’s Spring 2026 report.

From a commercial perspective, open models can significantly reduce the cost of deploying AI. Unlike closed models, which often involve high API fees and limited cloud-based deployment options, open models can be deployed privately on local infrastructure.

Moreover, as open-weight models approach the performance of closed-source rivals while being offered at a fraction of the cost and with greater deployment flexibility, closed-model providers are likely to face pressure on both pricing power and profit margins. A large-scale shift by developers toward open-weight solutions would mean that closed-source companies could not only lose direct API revenue, but also miss out on opportunities to build developer ecosystems and create a data flywheel.

For example, both DeepSeek V4 Pro, an open-source model, and SpaceXAI’s Grok 4.6, a closed model, which were released on the same day, rank among the top tier globally in terms of performance according to third-party benchmarks. However, DeepSeek maintains a significant price advantage, costing less than one-seventh of Grok 4.6. Although Grok 4.6 was released as a closed-source model, based on the US AI developer’s previous release pattern, Grok 4.6 could potentially be opened up after several rounds of model iterations. The same strategy would likely apply to the upcoming 4.7.

Competition in the open-source model market is expected to intensify further. In a recent open letter titled “The Future is Open,” Meta Chief Executive Officer Mark Zuckerberg said that stringent US restrictions on data use and model training are constraining the development of domestic AI labs, putting them at a significant competitive disadvantage relative to overseas teams. Simply banning foreign open-source models, he said, would not address the underlying issues.

Editor: Kim Taylor