Alibaba wants to charge the biggest users of its ‘open’ AI model

Alibaba is trying to make its free AI model pay. According to people familiar with the plan, the company intends to charge the biggest commercial users of its next open-source Qwen model, a shift that is subtle but significant.

The model itself would stay open-weight, so its underlying settings could still be freely downloaded and run by anyone who wanted them, yet the largest users, the ones building serious commercial products on top of it, would now have to negotiate a commercial agreement first.

That marks a break from how things worked before, since Alibaba has generally let companies run its open models on their own infrastructure without paying, using openness to win over developers and spread its technology. What makes monetisation tempting is Qwen’s success.

Alibaba has pushed the family hard, unveiling Qwen3.8-Max as its most capable model and closing the gap on the largest rivals, and the models have earned real standing along the way.

The company has claimed Qwen ranks as the world’s number-two open-weight model, which gives it the kind of reach that makes charging heavy users plausible.

The template comes from a rival. Alibaba is mirroring Moonshot, whose Kimi K3 model asks partners for revenue sharing once they cross a threshold, effectively taxing the biggest beneficiaries of its free release.

Those terms are specific, too: the arrangement can require up to 30% revenue sharing for partners with more than $20m in annual sales, a real cost for large deployers while staying free for everyone else.

Alibaba’s own terms are less settled, because the company has not yet fixed its revenue-share percentage, and the details are reportedly still being worked out as it prepares the next release.

The logic will be familiar to anyone who has followed software. It is essentially a freemium model, free for the many and paid for the few who make serious money from it, adapted here to the economics of large language models.

It also answers a hard question, since open models are expensive to train but generate no direct revenue when given away, and Chinese labs have been hunting for a way to fund the next round of development.

That search is unfolding against a fierce backdrop. A brutal price war has raged among China’s AI firms, with DeepSeek making a 75% discount permanent and squeezing everyone’s ability to charge for access.

The compute bills, meanwhile, are enormous. Training these models takes vast fleets of chips, and Moonshot reportedly used 20,000 Nvidia chips from Alibaba’s cloud, so someone eventually has to pay for that hardware.

The move also reframes what “open” means. If the biggest users must sign a contract, the label describes the licence more than the price, which blurs the line between open source and a commercial product.

There is a geopolitical edge to it as well, because China’s labs have leaned on openness partly to spread their models worldwide while US leaders kept their best systems closed, and charging heavy users risks blunting that advantage.

Even so, developers may not mind the trade. Smaller teams keep free access, while large enterprises are used to paying for support, tuning and reliability, so a paid tier for heavy commercial use may feel reasonable rather than a betrayal. The bigger test is whether openness can fund itself at all.

If revenue sharing works, it offers a path for open models to keep pace with the enormous budgets behind closed ones, and it does so without simply giving the technology away for nothing in return.

For Alibaba, then, this is a bet that reach can eventually be converted into revenue. Having spent heavily to make Qwen popular in the first place, it now wants to collect from the companies that have gone on to build real businesses on top of it.