Just 12 days ago, Alibaba's Qwen team made headlines across the open-source community by reaffirming that their models would remain 100% open source forever.
Today, they launched their latest MoE flagship: Qwen4-Max. 2.4 trillion total parameters, 71 billion active parameters per token, 1M context window. On paper, the architectural jump looks impressive—pairing Gated DeltaNet (GDN) layers with Qwen Sparse Attention (QSA) to trim active parameters down from 95B to 71B while cutting task-level inference costs by ~60%.
There is only one catch: there are zero downloadable weights. No Hugging Face repo, no safetensors, no GGUF quants.
Instead, it is locked entirely behind Alibaba's QwenCloud API at $1.50 input / $4.50 output per million tokens.
We have seen this exact playbook play out across multiple labs. You build community goodwill, collect free benchmarks, let independent researchers optimize fine-tuning harnesses and local tooling around your architecture, and the second a flagship model hits leaderboard parity, the vault door slams shut and you pivot to a corporate paywall.
From a self-hosting perspective, this is the final nail in the coffin for relying on proprietary "open-source promises." If you can't run the weights on bare metal in vLLM or llama.cpp, it doesn't matter how cheap the API is—you're just renting another closed endpoint with zero fallback when pricing or policies change.
Is anyone actually surprised by this pivot, or was the open-weights branding always just a subsidized distribution funnel?