Alibaba plans to introduce revenue-sharing phrases for some business customers of its subsequent Qwen open-weight AI mannequin, Reuters reported, citing two individuals aware of the corporate’s plans.
The association would require bigger corporations that generate income from providing the mannequin as a service to succeed in a business settlement with Alibaba. The precise revenue-sharing charge has not been finalised, the sources stated.
Alibaba is predicted to introduce the measure with its subsequent open-source mannequin. The corporate has beforehand charged builders to entry fashions hosted by means of its cloud platform, whereas typically permitting clients to deploy its open-source fashions in their very own information centres with out paying licensing charges.
The proposed association would differ from the licensing used for the present Qwen3 open-weight fashions, which Alibaba says are launched beneath the Apache 2.0 licence. The licence permits business use, modification, and redistribution topic to its situations.
Open-source versus open-weight
Open-weight fashions make their educated parameters accessible for obtain, however that doesn’t essentially imply each a part of the AI system is open or that each one types of business use are unrestricted.
Beneath the Open Supply Initiative’s Open Supply AI Definition, an open-source AI system ought to enable customers to make use of, examine, modify, and share it for any function with out looking for permission. OSI’s definition additionally requires entry to details about coaching information, related code, and mannequin parameters.
Alibaba and different Chinese language AI builders have launched giant fashions with downloadable weights. OpenAI, Anthropic, and Google, by comparability, primarily distribute their fundamental business fashions by means of closed techniques and hosted companies.
Alibaba’s deliberate phrases resemble the licensing mannequin adopted by Chinese language AI developer Moonshot for Kimi K3, an open-weight mannequin launched final month. Its licence consists of separate situations for corporations working Mannequin-as-a-Service companies above sure income thresholds.
Beneath Kimi K3’s printed licence, an organization working such a service should attain a separate settlement with Moonshot when the mixed income of the corporate and its associates exceeds $20 million throughout any consecutive 12-month interval. The availability applies to business use of Kimi K3 and spinoff fashions.
The licence comprises one other requirement for big consumer-facing deployments. Industrial merchandise exceeding both 100 million month-to-month energetic customers or $20 million in month-to-month income should prominently show the Kimi K3 identify, with exemptions masking inside use and companies supplied by means of Moonshot or licensed inference companions.
Based on two individuals aware of Moonshot’s business preparations, these agreements can embody income sharing. One supply stated Moonshot can require companions to share as much as 30% of the income concerned.
Chinese language IT companies firm Chinasoft Worldwide disclosed a revenue-sharing settlement with Moonshot in a regulatory submitting final month. It didn’t disclose the share concerned.
DigitalOcean Holdings can also be among the many corporations providing Kimi K3 and different Chinese language fashions. Chief Govt Paddy Srinivasan confirmed that DigitalOcean has a business settlement with Moonshot however declined to offer particulars.
Srinivasan described the strategy as an open-source “freemium” mannequin, the place corporations can entry software program at little or no preliminary price earlier than paying for larger-scale business use, technical companies, or earlier entry to future releases.
The price of operating open fashions at scale
Firms can obtain an open-weight mannequin with out paying for entry to an API, however giant fashions nonetheless require substantial computing infrastructure when deployed at scale.
Kimi K3 comprises 2.8 trillion whole parameters and 104 billion activated parameters, in accordance with Moonshot. Its mixture-of-experts structure consists of 896 specialists, with 16 chosen for every token.
The mannequin’s dimension nonetheless locations substantial {hardware} necessities on operators. Moonshot briefly stopped accepting new Kimi K3 subscriptions in July after saying utilization had positioned stress on its accessible GPUs, whereas Reuters reported that comparatively few customers have been anticipated to self-host a mannequin of that scale due to the infrastructure required.
Alibaba is utilizing an analogous architectural strategy with Qwen3.8-Max. The mannequin comprises about 2.4 trillion parameters however prompts round 95 billion parameters for every request, in accordance with Reuters.
Moonshot says its mixture-of-experts design improves scaling effectivity by activating solely a subset of the mannequin’s specialists for every token, somewhat than the complete mannequin.
Cloud suppliers can cost for internet hosting and inference, whereas AI infrastructure corporations can generate income from deployment and optimisation companies.
Dan Fu, vice chairman of kernels at Collectively AI, stated corporations offering AI companies can differentiate their choices by means of areas resembling extra environment friendly token use and deployment optimisation.
“On the software layer, there’s worth on the market for a way you utilize it, the way you truly get the fashions and the tokens to do one thing helpful,” Fu stated.
Mannequin growth presents a separate price problem. Analysis involving Epoch AI and Stanford researchers estimated that the price of probably the most compute-intensive coaching runs had risen by about 2.4 instances a yr since 2016, whereas Stanford’s 2025 AI Index discovered that the worth of accessing fashions at a given functionality stage had fallen sharply.
At printed API costs on the time of launch, Kimi K3 was priced at about one-third of Anthropic’s Fable mannequin based mostly on listed enter and output token charges. Pricing is just one a part of the deployment price, notably for corporations operating fashions on devoted infrastructure or dealing with excessive volumes of requests.
These prices sit alongside the licensing preparations being examined by mannequin builders. Alibaba already expenses builders that entry Qwen by means of Alibaba Cloud. The proposed association would additionally enable it to gather income from some corporations deploying Qwen independently on their very own infrastructure or by means of third-party companies.
Moonshot has already hooked up business situations to Kimi K3 whereas preserving its mannequin weights accessible for obtain. DigitalOcean and Chinasoft Worldwide have each disclosed business preparations with Moonshot, though the monetary phrases haven’t been made public.
The business preparations are creating alongside wider tensions between China and the US over AI know-how. The White Home has accused Moonshot of utilizing know-how taken from Anthropic whereas creating its fashions, an allegation Chinese language officers have rejected.
Curiosity in releasing fashions with downloadable weights shouldn’t be restricted to Chinese language builders. Considering Machines Lab, the San Francisco AI firm based by former OpenAI Chief Know-how Officer Mira Murati, launched its first open-source mannequin final month.
Lin Qiao, chief government and co-founder of Fireworks AI, stated there was no elementary technical barrier stopping US builders from releasing extra succesful open-source fashions. Fireworks AI works with fashions from builders together with Moonshot, though Qiao declined to debate its business preparations.
Alibaba has not publicly introduced the ultimate licence for its subsequent Qwen mannequin or the revenue-sharing proportion it plans to hunt from giant business customers.
(Picture by: Alibaba)
See additionally: Alibaba, DeepSeek push China’s AI mannequin race in direction of decrease prices
Wish to be taught extra about AI and massive information from trade leaders? Try AI & Big Data Expo happening in Amsterdam, California, and London. The excellent occasion is a part of TechEx and is co-located with different main know-how occasions together with the Cyber Security & Cloud Expo. Click on here for extra data.
AI Information is powered by TechForge Media. Discover different upcoming enterprise know-how occasions and webinars here.
