Открытые ИИ-модели Китая против OpenAI и Anthropic: чем берут DeepSeek и Qwen

Chinese AI labs led by DeepSeek and Alibaba with its Qwen family are, in July 2026, increasingly positioning their open-source models as a working replacement for the closed flagship models of OpenAI and Anthropic — against the backdrop of access to Western frontier models becoming more expensive and more restrictive. Wired reports this.

What's wrong with closed models

Access to OpenAI's and Anthropic's leading models has become noticeably more restricted in 2026: higher prices, tighter request limits, and more conditions around regions and moderation. According to Wired's assessment, this is exactly what has pushed developers to look for alternatives that can be downloaded, deployed on their own servers, and that don't depend on someone else's API and its policies.

For many teams, a closed subscription-based model is also an ongoing risk: the vendor can raise the price, change usage rules, restrict access by geography, or pull the model from the shelf altogether. Open weights remove that risk — a model that's already been downloaded can't be "recalled," and that undermines the very business logic of subscription-based access.

What the Chinese labs bring to the table

Chinese labs are betting on open weights and permissive licenses. DeepSeek, Alibaba (Qwen), Moonshot AI (Kimi), and Zhipu AI (GLM) are releasing models that can be run locally, fine-tuned for a specific task, and embedded into a product without calling out to an external server. Some of them are released under permissive licenses like Apache 2.0, which allows them to be used in commercial products.

Wired frames their pitch this way:

Chinese labs are promoting their open-source alternatives as "stable, affordable, and increasingly capable." —

Wired

• DeepSeek — a lab from Hangzhou known for open reasoning models
• Qwen — Alibaba's family of open models
• Kimi — Moonshot AI's models, with a focus on long context
• GLM — Zhipu AI's open models
• The common trump card — open weights and licenses that allow commercial use

The key difference from Silicon Valley's approach: the model is handed over in full, rather than through a paid API. It can be inspected, forked, and kept entirely under one's own control.

Why this matters for developers

Open models remove dependence on a single vendor: the team itself controls cost, data privacy, and access stability. Data doesn't leave for someone else's cloud, and inference cost is determined by one's own hardware rather than a vendor's price list.

For companies outside the US, this is also a matter of technological sovereignty — not being cut off from a tool due to sanctions, export restrictions, or a vendor's sudden policy change. According to Wired, in 2026 the main argument in favor of Chinese open models is becoming predictability of access, not just price — and as these models become "increasingly capable," the quality gap with closed flagship models is narrowing as well.

What this means

The industry balance is shifting from closed APIs toward open weights. If Silicon Valley's "playbook" has for years been built on proprietary, subscription-based models, Chinese labs are demonstrating the opposite strategy: openness itself is becoming a way to capture the market and standardize development around their models. For OpenAI and Anthropic, this means that the closedness of frontier models is gradually turning from an advantage into a vulnerability.