Alibaba’s Qwen team has released an early look at what will become Qwen4. The preview, called Qwen3.8-Flash-Next, packs 125 billion total parameters but switches on just 6 billion of them to generate any given token. The interesting part isn’t the ratio itself; it’s the mechanism producing it, and a licensing detail that could keep the model outside the EU’s lightest regulatory tier.
Cutting active parameters usually means adding more experts to a mixture-of-experts model and routing each token to a handful of them. Qwen’s engineers took a different path this time. They attached a 51 billion parameter block that functions as a lookup table, keyed to two and three character text fragments instead of an expert-routing network. The team says querying that table costs less compute and offloads more easily onto accelerators that are constrained on memory, the kind of hardware limit export controls have made routine for Chinese labs building at the frontier.
A 6-billion-of-125-billion activation ratio has a concrete consequence for who can run this thing. It means inference can plausibly happen on a single high-memory accelerator, not the multi-node cluster a dense model of comparable size would demand. That lowers the entry cost for any team wanting to self-host rather than pay a frontier lab’s API rate, which is the commercial bet behind the architecture as much as any benchmark score.
Alibaba frames the redesign around inference economics, not raw capability. Its point of comparison is its own prior release, Qwen3.7-Plus, which holds 397 billion parameters and activates 17 billion of them, leaving the new preview running at close to one third of that predecessor’s active-parameter footprint.
Four design choices produce the smaller footprint. Three follow well established engineering patterns: a sparse-attention mechanism that scores contiguous blocks rather than scoring tokens one at a time, a residual-gating layer that filters what information passes between layers, and a training recipe that entirely skips batch-size warmup. The fragment-indexed lookup table is the genuine departure from that list.
Alibaba disclosed one limitation candidly. The model card states that GPT-4o, not the benchmark’s own grading model, scored the preview’s result on Humanity’s Last Exam, an admission that argues for discounting the figure rather than trusting it more.
Every performance number attached to this release still comes from Alibaba’s own testing harness. The Next Web has pointed out before that Alibaba billed Qwen3.8 as second only to one rival worldwide, a ranking it never backed with outside verification, and the same caveat applies here.
The licence is where this stops being a purely technical story. The weights are posted to Hugging Face carrying a qwen-community licence. In an August 7 report, The Next Web said Alibaba intends to bill its biggest paying customers for use of the model. Under the EU AI Act, Article 53(2) waives two documentation obligations for models released under a genuinely free and open-source licence, while Recital 103 excludes anything monetised or provided at a price from that exemption. A licence structured to bill enterprise customers is close to the exact scenario Recital 103 describes.
That distinction already has real deployments riding on it. Thomson Reuters has built a product on Qwen, and any European company doing the same inherits whatever regulatory status the licence eventually lands on. Legal and compliance teams evaluating a Qwen-based deployment for EU customers should not assume the open-source documentation exemption applies until Alibaba clarifies the commercial terms attached to the qwen-community licence.
Reported by The Next Web on 26 August 2026.
