Why Is China Releasing Open Source AI Models at the Very Top of the Market?

Alibaba announced on 3 August 2026 that it had released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model, and that it would open-source the weights alongside a smaller sibling, Qwen3.8-27B. Bloomberg and the South China Morning Post both framed the release as a capability milestone. The more interesting story is the pricing sheet that came with it, because it explains a strategy that has been building across Chinese AI labs for two years: give away, or nearly give away, the layer where American labs make their money.

A frontier model priced like a commodity

Qwen3.8-Max is available through Alibaba's API at $2 per million input tokens and $6 per million output tokens, according to reporting by The Information. That is roughly a third of the cost of Moonshot's rival Kimi K3, a 2.8-trillion-parameter model priced at $3 and $15 respectively. Both models support a 1-million-token context window and native multimodal input, putting them on paper alongside the best commercial systems from OpenAI and Google. The difference is that Qwen3.8-Max's weights are being published, meaning anyone with sufficient hardware can run it without paying Alibaba anything at all.

Why give away the expensive part

Training a model of this size costs hundreds of millions of dollars in compute alone. Open-sourcing it looks, at first glance, like giving away the asset that justified the spend. But the calculation changes once you separate the value of training a frontier model from the value of selling access to one. Chinese labs are increasingly betting that the API-access business — the business US labs depend on for revenue — is the part worth destroying, not preserving.

Open-weighting the frontier does not give away the advantage of having trained it first. It removes the advantage of being the only one who can serve it.

If a 2.4-trillion-parameter model is available for free download, any cloud provider, integrator or state-backed lab worldwide can host it and undercut proprietary API pricing. Reuters has reported separately that DeepSeek's newest release now benchmarks as the cheapest well-known model to run, a reputation the company has cultivated deliberately since its original January 2025 release rattled the market. Qwen and Kimi are following the same script at a larger scale.

Where the money actually sits

• Alibaba and Tencent sell the cloud compute that running a 2.4-trillion-parameter model at scale requires, regardless of who trained it.
• Distribution into China's domestic cloud, e-commerce and enterprise software stacks captures value independent of API margin.
• A widely adopted open-weight model becomes the default building block for downstream products, entrenching an ecosystem rather than a single subscription.
• Undercutting US API pricing slows the revenue growth that Western labs need to justify further training runs.

The strategic read

American frontier labs have built business models around charging a premium for access to models they alone can run economically. Open-weighting a model of Qwen3.8-Max's scale attacks that premium directly: it does not need to be better than GPT-class or Claude-class systems, only good enough and free enough that customers stop paying for the alternative. Commoditising the layer where a competitor earns margin, while retaining the compute and distribution layers underneath it, is a strategy with obvious precedent in software history, and Chinese labs are running it in public.

What to watch

Watch whether US labs respond with their own price cuts or with tighter licensing on their most capable models, and watch adoption figures for Qwen3.8-Max and Kimi K3 outside China over the next two quarters. If enterprises in Europe and Southeast Asia begin quietly routing production workloads through open-weighted Chinese models because the economics are simply better, the pricing weapon will have worked regardless of which lab's model tests marginally higher on any given benchmark.