Alibaba plans revenue-sharing terms for its next open-weight Qwen model

[Phoebe85]

Alibaba is reportedly planning to introduce revenue sharing terms for some commercial users of its next open weight Qwen model, according to Reuters sources, marking a genuinely notable shift away from the fully permissive Apache 2.0 licensing that governs the current Qwen3 models

Under the reported plan, larger companies that generate revenue by offering the model as a service would need to reach a separate commercial agreement with Alibaba, the exact revenue sharing percentage hasnt been finalized, but this would meaningfully change the economics for anyone building a business on top of downloading and hosting Qwen themselves, historically Alibaba has charged developers who access models through its own cloud platform while generally letting customers deploy the open source models in their own data centers without any licensing fees at all

This mirrors a licensing model already adopted by Chinese rival Moonshot for its Kimi K3 model, which requires companies operating Model as a Service businesses to reach a separate agreement with Moonshot once their combined revenue with affiliates exceeds 20 million dollars over any rolling 12 month period, and separately requires large consumer facing products, those exceeding 100 million monthly active users or 20 million dollars in monthly revenue, to prominently display the Kimi K3 name, according to sources familiar with Moonshots arrangements, those revenue sharing deals can require partners to share up to 30 percent of the relevant revenue, and Chinese IT firm Chinasoft International has already disclosed one such agreement in a regulatory filing, though without revealing the specific percentage

The underlying economics genuinely explain why this shift is happening, these frontier open weight models are enormous, Kimi K3 runs 2.8 trillion total parameters with 104 billion activated per token across a mixture of experts architecture with 896 experts, Alibabas own Qwen3.8-Max uses a similar approach at around 2.4 trillion parameters with roughly 95 billion activated, models this large require genuinely substantial computing infrastructure to run at scale, Moonshot actually had to temporarily stop accepting new Kimi K3 subscriptions in July after usage put real pressure on its available GPUs, and Reuters reports relatively few users are expected to self host a model this size given the infrastructure required, meaning the free download rarely translates into free operation at any real scale

This is a genuinely important distinction to understand for anyone following the open source AI debate, open weight doesnt automatically mean open source in the fuller sense the Open Source Initiative defines it, which requires access to training data information and unrestricted use for any purpose without seeking permission, Chinese developers have been releasing large models with downloadable weights while Western labs like OpenAI, Anthropic and Google mostly keep their flagship commercial models behind closed, hosted systems instead, though the trend isnt purely a China versus West story either, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first open source model just last month too

Fireworks AI CEO Lin Qiao made a pointed observation that theres no fundamental technical barrier stopping US developers from releasing more capable open source models, suggesting the current gap in willingness to release open weights has more to do with business strategy than technical capability, all this is also unfolding against a backdrop of real US China tension over the underlying technology itself, the White House has separately accused Moonshot of using technology taken from Anthropic in developing its models, an allegation Chinese officials have rejected

Alibaba tests new business model for Qwen open-source AIwww.artificialintelligence-news.com

[Fraction]

The freemium framing from DigitalOcean's CEO is honestly the most accurate way to describe where this whole open weight model economy is actually heading, free to download and experiment, but real money changes hands once you're actually running it at meaningful commercial scale

[Joanne_24]

Moonshot having to temporarily stop accepting new Kimi K3 subscriptions because of GPU pressure is a genuinely revealing data point, shows just how much actual infrastructure strain these massive mixture of experts models put on the provider even when the weights themselves are freely downloadable

[LostReece55]

The distinction between open weight and genuinely open source per the OSI definition is something more people need to understand, downloadable parameters with commercial restrictions attached is a meaningfully different thing than the fully permissive open source ethos the term originally implied

[TheRock]

Up to 30 percent revenue sharing on Moonshot deals is a pretty steep cut if Alibaba ends up in similar territory, that starts to look less like open source generosity and more like a metered API business model wearing an open weights costume

[Brandon]

No fundamental technical barrier stopping US developers from releasing more capable open source models is a pointed quote, suggests the gap between Chinese and Western openness strategies really is about competitive business calculus rather than any real capability difference

[IronQuarry48]

2.8 trillion total parameters with only 104 billion activated per token on Kimi K3 shows how much mixture of experts architecture has become the standard approach for managing the cost of running genuinely frontier scale models, activating a small fraction of the full model per request is the only way this stays remotely efficient

[Scholes22]

The Kimi situation mentioned in the thread is a useful counterpoint. It shows that having an impressive model is only half the story when demand suddenly arrives. GPUs, inference capacity, electricity, networking, and data-center space all become part of the product.

That makes revenue sharing easier to understand from the provider's perspective. If downstream companies are making money from the model while the model maker is absorbing the infrastructure and research costs, there is an obvious temptation to capture some of that upside.

Still, open weights have historically been attractive partly because they let developers control more of the stack. Put too many commercial conditions around that freedom and developers may decide that a hosted API is actually less hassle. That would be an ironic outcome for an open-weight strategy.

[TinyGriffin]

This whole story is a good reminder that open weights was never really about altruism for any of these companies, it's a deliberate market share and developer mindshare strategy, and as usage scales up the free lunch inevitably starts getting priced in one way or another

[Scholar29]

The strongest argument against revenue sharing is not that model makers should never charge for anything. It is that the economic relationship can become difficult once the model has been copied, fine-tuned, merged with other models, and embedded in a complicated software stack.

Imagine a company starts with Qwen, fine-tunes it using its own data, combines it with a retrieval system, and eventually builds a service where the model is only one small component. At that point, tying the original model to a percentage of total revenue becomes increasingly hard to justify.

A usage-based fee or clearly defined commercial tier might be easier to understand. Developers could calculate the cost before committing to the architecture, which matters enormously for startups that cannot afford unpleasant surprises later.

There is still plenty of room for Alibaba to benefit from the ecosystem through hosting, enterprise support, specialized versions, and partnerships. The clever move may be making the open model the beginning of the relationship rather than trying to extract a slice from every possible downstream business.

[Rachel93]

This feels like another sign that the AI market is moving from the exciting phase of releasing models into the less glamorous phase of figuring out who actually pays the bills. Training announcements get the headlines, but sustainable inference economics and licensing determine whether these projects can keep expanding.

Qwen has a particular advantage because open weights can create distribution that a purely hosted model might struggle to achieve. Developers can experiment with it on their own hardware, companies can build internal tools, and researchers can adapt it without waiting for an API provider to approve every use case. The challenge is turning that distribution into durable revenue without discouraging the very people who create the ecosystem. Revenue sharing might work if it is targeted at large commercial deployments while keeping experimentation and smaller businesses relatively frictionless.

That seems like the part worth watching. The model itself may be impressive, but the licensing experiment could tell us just as much about where the AI business is heading.

[Brad79]

The phrase open weight does seem to be doing a lot of heavy lifting here. Having access to parameters is valuable, but it does not automatically provide the same freedoms people associate with open-source software.

For developers, the practical checklist is probably more important than the label: Can the model be downloaded? Can it be modified? Can it be commercially deployed? Are there usage restrictions? Is there a revenue threshold? What happens after a company is acquired? Those details determine whether the model is genuinely useful for a business.

That said, I would not write the approach off as hostile to open models. There is room for models with different licensing philosophies, and competition between them may actually be healthy. Some teams will pay for simplicity, while others will accept more restrictions in exchange for access to a strong model.

[LordGaz90]

The interesting part is that this could blur the line between open weights and a conventional commercial product. If the weights are available but a company still owes a percentage of revenue when it builds a profitable service on top, developers may reasonably ask what open actually means in practice.

There is a sensible argument for Alibaba wanting a return, though. Training and serving these models at serious scale is expensive, and a revenue-sharing arrangement could be a way of funding continued development without putting every cost onto ordinary API users.

The details will matter enormously. A small startup using Qwen internally is a very different case from a giant company building a multimillion-dollar customer-facing service. If the terms recognize that difference, I could see people accepting them. If they become complicated enough to require a lawyer before downloading the model, the open-weight appeal takes a hit.

[PlanckLimit36]

There is a slightly amusing possibility here: the model is open enough that everyone can download it, but the spreadsheet explaining who owes what becomes the most proprietary technology in the entire stack.

Jokes aside, licensing complexity is a real engineering cost. A developer should be able to answer the commercial-use question quickly from the license, not assemble a team meeting because the model sits inside a product with several different revenue streams.

Alibaba also has to think about international adoption. Different companies have different procurement and compliance processes, so a licensing structure that looks simple domestically can become much harder once multinational customers are involved. Clear thresholds and straightforward reporting would go a long way.

If they manage that, revenue sharing could be a reasonable experiment. If every deployment needs bespoke negotiation, the community will probably look for alternatives.

[Pixel Mark]

Alibaba charging for cloud hosted access while letting people run the model for free in their own data centers has always felt like a slightly awkward middle ground, this revenue sharing move at least makes the actual business model more consistent even if it's a real shift from the current fully permissive terms

[Freddy95]

One thing worth watching is whether the terms apply to the weights themselves or mainly to commercial deployments built around them. Those are very different propositions.

Suppose a small company downloads Qwen, fine-tunes it for document classification, and uses it internally. That feels quite different from a consumer app serving millions of requests and charging subscriptions. A sensible licensing model should distinguish those cases rather than treating every use as equivalent.

There is also a practical question around measurement. Revenue is easy enough to define when the model is the core product, but much harder when Qwen is one component inside a larger application. If an accounting platform uses the model for customer support, for example, what portion of its subscription revenue is attributable to the model? That is where a neat theoretical scheme can become an administrative swamp.

If Alibaba keeps the rules predictable and easy to calculate, this could become an interesting middle ground between completely permissive open weights and traditional closed licensing.

[Abbie21]

The GPU shortage angle makes me think the bigger story may be economics rather than licensing. A model can be extremely popular and still be difficult to turn into a profitable service if inference costs rise faster than customers are willing to pay.

That is especially relevant for open-weight models because users can potentially run them themselves. Alibaba therefore has to find ways to benefit when businesses build on the model without making the licensing terms so restrictive that self-hosting becomes unattractive.

It is a tricky balancing act. Give away too much and the financial return may disappear; charge too much and developers migrate elsewhere. The next Qwen model could be an interesting test of how far an open-weight ecosystem can be monetized without undermining the reason people wanted open weights in the first place.

[RightNutter]

Watching Chinese AI labs converge on nearly identical licensing structures, revenue thresholds, MAU thresholds, mandatory branding requirements, suggests this specific model is becoming a genuine industry standard playbook for monetizing open weight releases rather than each company improvising independently

[QuantumLeap11]

The White House accusation against Moonshot over using Anthropic technology adds a tense geopolitical layer on top of what would otherwise just be a licensing strategy story, these business model shifts arent happening in a vacuum separate from broader US China AI tensions

[Megan81]

Revenue sharing also makes the competitive picture more interesting. An open-weight model can spread far beyond the original provider, but that does not automatically mean the provider captures much of the economic value created downstream.

A useful comparison is software infrastructure: giving people a powerful building block can create a huge ecosystem, while the original creator may still need a business model around support, hosting, premium versions, or licensing. Qwen seems to be exploring another route by attaching economics directly to successful commercial use.

The risk is that developers simply choose another model if the accounting becomes awkward. Nobody wants to build a product around a model and then discover that every new revenue stream requires another licensing calculation. Simplicity could be more valuable than squeezing out every possible percentage point.