The More Alibaba Cloud (阿里云) Makes, the More Dangerous the Open Source Ecosystem Becomes

A closed-door turn at the multimodal frontier

Alibaba (阿里) on March 30 unveiled Qwen3.5‑Omni, a multimodal model that handles text, images, audio and video and supports 113 languages for speech recognition. Alibaba said the model surpasses Google’s Gemini 3.1 Pro on several audio benchmarks; it has been reported that the company is marketing three variants — Plus, Flash and Light — that are API‑only. That matters because the prior Omni generation was released under Apache 2.0 with freely downloadable weights. Why close what you once opened? For developers who relied on local deployment, fine‑tuning and low‑cost experimentation, the change is immediate and painful.

Business logic: tokens, APIs and cash pressure

The move follows an internal re‑orientation toward token‑based monetization. It has been reported that a new ATH organisation — Alibaba Token Hub — was created to centralise token production, distribution and monetisation, and that internal disagreements over open‑first versus token‑first strategies contributed to staff departures. The pivot is also financial. Alibaba Cloud’s AI businesses have posted rapid revenue growth, but group profitability and free cash flow have been under pressure after heavy capex on chips, datacentre compute and model training. Management says MaaS (model‑as‑a‑service) is set to be a major revenue stream; when your largest product is an API that consumes compute with every call, the incentive to fence off frontier capabilities is strong. Against the backdrop of U.S. export controls and intensifying U.S.‑China tech competition, Chinese cloud providers also face strategic reasons to control how advanced models are distributed.

What this means for open source and innovation

There is a paradox. Open releases helped Qwen become a developer brand — millions of downloads and thousands of derivatives created a funnel of users who later migrated to paid APIs. Reportedly, much of recent token consumption growth was seeded by that open ecosystem. But locking the most commercially valuable multimodal capabilities behind APIs risks starving the open community of the freshest architectures, largest compute allocations and top research talent. To be clear: base Qwen models remain open, so this is not an outright reversal. Still, as frontier capabilities are systematically channelled into closed, billable products, the balance between public‑good open research and platform monetisation shifts — and the global open‑source AI ecosystem may be the one to lose most.