The new power move: give state-of-the-art AI away
Open-source AI models are freely available machine learning systems whose underlying weights and licenses allow developers to download, run, and modify them locally or in the cloud, reducing dependency on closed, pay-per-call APIs while shifting power back to teams that can integrate and operate the models effectively over time.
Alibaba and Tencent have decided that the strongest play in the AI race is not another closed API, but free open-weight models that rival premium systems. Alibaba has released Qwen3.8-Max, its most capable model to date, as open weights, the first time a Max-class Qwen model has been opened this way. Tencent is expanding global access to Hy3, an open-source large language model, through tools like WorkBuddy, Tencent Design Miora, and Tencent Cloud TokenHub. The message is blunt: if you are still fully tied to expensive third-party APIs, you are leaving strategic leverage on the table.
Inside Qwen Max and Hy3: free AI downloads that are not toys
Qwen3.8-Max and the Tencent Hy3 model are not hobbyist experiments; they sit squarely in the top tier of open-source AI models. Qwen3.8-Max runs 2.4 trillion parameters in total, with 95 billion active at any moment, using a Mixture-of-Experts design to keep compute requirements manageable. The weights will be available as free AI downloads on major model hubs, marking the first time Alibaba has opened a Max-scale Qwen model. In parallel, Hy3 is a 295B-parameter Mixture-of-Experts model with 21 billion active parameters, a 256K-token context window, and an Apache 2.0 license. That combination of scale, long context, and permissive licensing makes both systems credible options for local AI deployment instead of purely remote consumption.
| Spec | Alibaba Qwen3.8-Max | Tencent Hy3 |
|---|---|---|
| Total parameters | 2.4T | 295B |
| Active parameters | 95B switched on | 21B active |
| Context window | Not stated | 256K tokens |
| License / access | Open weights via model hubs | Apache 2.0, open-source |
What developers and enterprises actually gain
The practical shift is control. Small businesses and labs with capable hardware can now run a state-of-the-art model without paying for massive datacenter-scale infrastructure. Qwen has already become the most self-hosted model globally, overtaking a leading Western open model, which shows how quickly open-weight adoption can snowball once credible options exist. Tencent’s Hy3 is explicitly designed for reasoning, agent workflows, coding, and other productivity tasks, and the company says developers can self-host it through mainstream inference frameworks. Enterprises that want to keep sensitive workloads in-house can move beyond pure API consumption and design their own local AI deployment strategies, from private copilots to custom agents wired into internal systems.
On the user side, Tencent is going even further: WorkBuddy users worldwide can use Hy3 for free until 31 August 2026, giving both individuals and companies a long runway to test and embed the model in daily workflows. Tencent reports that, in internal evaluations using WorkBuddy, Hy3 achieved a task success rate of more than 90% and reduced average task completion time by 34% compared with its previous model iteration. That is the kind of concrete productivity gain that turns experiments into default tools.
Strategic timing: export controls, shifting tokens, and open-weight geopolitics
These free AI downloads are not arriving in a vacuum. Interest in Chinese open-source AI models from overseas developers has been rising, even as enterprise adoption in some markets is still constrained by data privacy, compliance, and security concerns. At the same time, Chinese open-weight models have gone from under 2% of tokens on a major routing platform in late 2024 to roughly 61% by mid-2026. One quotable way to frame this shift is: "Chinese open-weight models went from under 2% of tokens on OpenRouter in late 2024 to roughly 61% by mid-2026". Add in export controls restricting certain American models and reports of potential limits on Chinese models going overseas, and open weights become both a distribution weapon and a hedge against political risk. Alibaba is candidly “losing on paper and winning on distribution” by betting that being good, free, and downloadable beats being slightly better but locked away.
The trade-offs: open source power, closed source polish
For all the excitement, developers should not romanticize open-source AI models as universal replacements for the best closed systems. On Alibaba’s own benchmark table, a leading American model (Fable 5) wins 15 of 31 text tests, another (GPT‑5.6 Sol) wins nine, and Qwen3.8-Max wins seven. On 12 coding tests, Qwen wins exactly one. That gap matters if your business lives or dies on code quality or complex reasoning speed. However, Qwen’s intelligence cost is described as much cheaper than at least one rival, which means that even if it needs more iterations, the total cost to reach a correct answer can still be attractive. Meanwhile, Hy3 is already targeted at reasoning, agents, and coding tasks, and Tencent Cloud is working with partners in South Korea and Japan to put it onto enterprise AI platforms. The rational strategy for enterprises is not ideological: keep closed models where they are clearly ahead, and aggressively move everything else to local AI deployment with open weights.
The conclusion is straightforward: open-weight heavyweights like Alibaba Qwen Max and the Tencent Hy3 model are enough to make pure API dependence look lazy. Developers should be planning hybrid stacks where open models handle most workload volume, closed models handle edge cases, and ownership of infrastructure and data stays firmly on their side. If you can download something close to the best for free, being second in benchmarks is not a weakness—it is a business model.
