For a while, parts of the AI community feared that Qwen might be moving away from open models. Qwen3.7 arrived as a hosted product without a comparable open-weight release, raising questions about whether Alibaba would continue supporting the local AI ecosystem.
Qwen3.8 provides a reassuring answer: Qwen is still releasing open weights—and this is a substantial update.
The release includes two very different models. At the top is Qwen3.8-2.4T-A95B, a massive mixture-of-experts model containing 2.4 trillion parameters, with approximately 95 billion activated for each token. This is the first time Qwen has released weights from its Max-class family.
However, the downloadable model is not identical to the hosted Qwen3.8-Max service. The open-weight version is text-only and has a native context window of 262,144 tokens, although Qwen says it can be extended to roughly one million. The hosted service adds vision input, non-thinking mode, a default one-million-token context window and built-in tools.
At nearly five terabytes in BF16, the 2.4T model is primarily relevant to cloud providers and well-equipped research labs—not home users.
The more practical release is Qwen3.8-27B. This dense 27-billion-parameter model supports text, images and video, has a native 262K context window and uses the permissive Apache 2.0 license. Quantized versions can run on high-end personal workstations, making it the model most local AI users are likely to try.
The same architecture, but better results
Interestingly, Qwen3.8-27B uses the same visible architecture as Qwen3.6-27B: 64 layers, a 5,120-dimensional hidden size and the same hybrid combination of Gated DeltaNet and conventional attention.
That does not mean the checkpoints or their behavior are identical. Instead, Qwen appears to have achieved most of the improvement through better training and post-training rather than redesigning the architecture.
According to Qwen’s evaluations, Qwen3.8-27B improves over Qwen3.6-27B from 53.5 to 61.7 on SWE-bench Pro, from 63.9 to 84.3 on OSWorld-Verified, and from 48.8 to 64.8 on WebArena-Verified.
These are vendor-reported results, and some evaluations use Qwen’s own benchmarks or testing harnesses, so independent verification will remain important. Nevertheless, the reported gains are impressive for a model using essentially the same underlying design.
The license may be as important as the model
For me, the licensing change is almost as interesting as the model itself.
Qwen3.8-27B remains under Apache 2.0, but the 2.4T model uses the custom Qwen3.8-Max License. It permits use, modification, distribution, fine-tuning and commercial deployment, but introduces conditions for very large businesses.
If a commercial product exceeds 100 million monthly active users or US$20 million in monthly revenue, the model’s name must be prominently displayed in its interface. This threshold triggers attribution—not a general licensing fee.
A separate Qwen license is required more narrowly when a company operates a qualifying model-as-a-service or AI work-assistant business and it, together with its affiliates, exceeds US$50 million in aggregate revenue during any consecutive 12-month period. The requirement does not apply to purely internal use that does not expose the model, its outputs or its capabilities to third parties.
This is not “open source” in the conventional Open Source Initiative sense, because the license places conditions on certain fields and scales of commercial use. Open-weight or source-available is therefore the more precise description.
But that does not make the experiment less significant.
Alibaba is attempting to offer the best of both worlds. Individuals and smaller businesses can run the model on their own infrastructure, maintaining sovereignty over their data and systems. At the same time, a large competitor cannot necessarily take Qwen’s work, turn it into a major competing model service or coding assistant, and generate substantial revenue without meeting additional conditions.
Perhaps this hybrid approach will become a template for future frontier models: broad access for local users, researchers and smaller businesses, combined with negotiated terms for companies commercializing the model at enormous scale.
I would not be surprised if other open-weight providers—or even traditionally closed model companies—explored similar arrangements over the coming months. The future may not be divided simply between completely open and completely closed models. A growing middle ground could allow public access without requiring model developers to give up all their commercial leverage.
Qwen3.8 is therefore good news for the open-model community on two fronts. The 27B release is compact, multimodal, commercially friendly and apparently much more capable than its predecessor. The Max release, meanwhile, tests a potentially important new business model for distributing frontier weights.
The most consequential innovation in Qwen3.8 may not only be how the models were trained, but also how Alibaba has chosen to share them.
Leave a Comment