Alibaba's Qwen team is set to release Qwen 3.8-Flash-Next on Wednesday, a 125-billion-parameter model that activates just 6 billion parameters per token. The team has framed the release as a preview of the next-generation Qwen 4 architecture, not a finished flagship.

According to the team's pre-release briefing, the model carries 125 billion total parameters with only 6 billion active per token. Hard benchmark scores have not been published, and the weights were not yet live on ModelScope, Alibaba's own open-source model hub, as of this writing.

There is no official information on the model beyond the announcement, but based on rumors, it will likely be a mixture-of-experts system — a design that splits the network into many specialized sub-models and lights up only the relevant ones for each task. A 125-billion-parameter model built this way would run with the compute bill of one that is just 6 billion parameters.

Parameters are essentially all the dials a model can tweak. The more parameters, the more capable a model is and the more computing power it requires. A mixture-of-experts design makes it possible for an extremely powerful model to activate only what it needs to produce the best output without wasting resources. The technique is an industry staple rather than a Qwen invention: Mistral's open-weight Mixtral models brought sparse mixture-of-experts designs to wide attention, DeepSeek's V3 runs on the same idea, and Qwen's own Qwen 3 line already includes MoE variants such as Qwen3-235B-A22B, which activates 22 billion of its 235 billion parameters per token. What this release previews is a new architecture generation, not mixture-of-experts itself.

Qwen 3.8 Flash Next is releasing Tomorrow. 125B paramters +51B N-gram and 6B active. Its based on the next generation Qwen 4 architecture. Qwen 4 is coming pic.twitter.com/rAcBZlNwKs

— AiBattle (@AiBattle_) August 25, 2026

Alibaba's Qwen team does describe the model as multimodal and built on the upcoming Qwen 4 architecture, and says it shipped the early build so developers can prepare for the full family.

Why the "3.8" isn't the news

Alibaba has put out a teaser, though the team itself calls it a preview. The plan is to ship the architecture improvements now, ahead of the complete Qwen 4 rollout. Hugging Face, where the weights also live, likewise describes it as "a preview of the Qwen 4 architecture."

Hard benchmarks haven't landed yet. Qwen hasn't published side-by-side scores against its own Qwen 3 line or Western rivals, so the 125 billion and 6 billion parameter figures are not verified, and its performance can only be speculated on.

China's open-weight cadence has been relentless. A mysterious free model, Ox Alpha, recently beat Anthropic's Fable on certain coding benchmarks, with no known builder behind it. Alibaba, DeepSeek, and Moonshot have all shipped capable weights that anyone can download, fine-tune, and run.

Open weights let developers build without sending data to a closed API, and they undercut the cost of hosted models. That's why an open 125-billion-parameter model with 6 billion active parameters matters: it puts near-frontier capability on commodity hardware.

But the numbers will have to wait. Qwen hadn't posted benchmark scores for the release as of this writing. Until they land, the concrete checkpoints are the weights going live on ModelScope and Hugging Face, independent benchmark runs from developers who download them, and whatever release timeline Qwen attaches to the full Qwen 4 family.