Alibaba ships Qwen 3.8's open weights, and the license splits the release in two

When Alibaba previewed Qwen 3.8-Max at a Shanghai conference in July, the loudest promise was also the vaguest: open weights, coming “soon.” On August 12 that promise resolved. Alibaba published the weights for its 2.4-trillion-parameter flagship on Hugging Face and ModelScope, and both NVIDIA and the vLLM project shipped Day-0 serving support the same day. The release is real. It is also not one thing: what Alibaba opened is two models under two very different licenses, and the gap between them is where a working team’s decision actually lands.

From “soon” to shipped

Two artifacts arrived within days of each other. The flagship, Qwen3.8-2.4T-A95B, is a fine-grained mixture-of-experts model with 2.4 trillion total parameters that activates 95 billion per token, the detail the July preview left blank. A smaller Qwen3.8-27B followed, including an FP8-quantized build that trims the memory footprint for single-node deployment. Both list a native context window of 262,144 tokens, extensible further, and NVIDIA measured the flagship at over 4,000 tokens per second per GPU and over 350 tokens per second per user on its GB300 NVL72 rack in FP8 precision, without additional tuning.

The activated-parameter figure matters more than the headline count. A 2.4-trillion-parameter dense model would be out of reach for almost everyone; a 95-billion-active sparse model is a different economic proposition, closer in inference cost to a mid-size dense model even though the full weights are enormous. That is what makes “open weights at this scale” more than a talking point: the compute you rent per token is set by the active path, not the total.

Two models, two licenses

The July preview flagged the license as the load-bearing question, and the answer is a split. The 27B ships under Apache 2.0, a permissive, commercially clean license: teams can self-host it, fine-tune it, and put it in a product without negotiating terms, and it accepts images and video alongside text. The 2.4-trillion flagship does not. It ships under a custom license Alibaba labels “qwen3.8-max,” it is text-only with multimodal input explicitly unsupported, and its model card states that reasoning mode is always on and cannot be disabled.

That last constraint is easy to skim past and expensive to ignore. A model that always thinks before answering spends tokens, latency, and money on every call, including the short ones where a plain answer would do. Paired with a non-standard license, the flagship is “open” in the sense that you can download the weights, not in the sense that you can drop it into any product on any terms. For most teams, the permissive 27B rather than the restricted flagship is the artifact that changes their options.

What it means for developers

The practical read is a fork. If the goal is a self-hosted model you can ship commercially without a license review, the multimodal 27B under Apache 2.0 is the one to test, and early interest from the local-inference community suggests teams are already doing exactly that. If the goal is the flagship’s capability, the first question is whether you can serve a 95-billion-active model at all: Day-0 support from NVIDIA and vLLM lowers the deployment friction, but the hardware bar is real, and running the Max tier is a different commitment than pulling a 27B onto a single node.

For the open-model race, the signal is that the frontier tier and the open tier keep converging. A vendor with Alibaba’s distribution putting a near-frontier-scale model into public weights, weeks after DeepSeek shipped V4 Flash under an MIT license, keeps the pressure on every provider pricing a mid-tier model behind a closed API.

What is still unverified

One thing has not changed since July: the capability ranking is still Alibaba’s own. The benchmark scores it published alongside the model are a vendor self-report, and no independent evaluation has yet reproduced the “second only to Fable 5” framing the preview leaned on. The difference now is that public weights make that claim testable in a way the preview did not, which is the healthiest part of this release.

Stackmaven’s catalog keeps Qwen on a watch footing. The signals that would move it: independent agentic and coding results outside Alibaba’s own table, evidence that the forced-reasoning flagship is affordable to run in a real agent loop, and adoption of the 27B in serious open-source stacks. Expect a reassessment on or around November 15.

Sources cited

1. Qwen3.8-2.4T-A95B model card (Alibaba, Hugging Face) huggingface.co
2. Qwen3.8-27B-FP8 model card (Alibaba, Hugging Face) huggingface.co
3. Serve Qwen3.8-2.4T-A95B on NVIDIA GB300 NVL72 (NVIDIA Technical Blog) developer.nvidia.com
4. Day 0 Support for Qwen3.8-2.4T-A95B (vLLM Blog) vllm.ai