Alibaba's Next Open Model Is a Preview of Qwen 4
Qwen3.8-Flash-Next is a 125-billion-parameter open-weight model that only uses six billion of them per word. It lands today, and it is built on the architecture behind the coming Qwen 4 family.
Alibaba put a countdown timer on its ModelScope platform on August 25 for Qwen3.8-Flash-Next, a new open-weight model releasing today, August 26. Open-weight means the model files themselves are published, so anyone can download and run them rather than only reaching the model through a paid interface. The interesting part is not the model but what it previews: Alibaba says it is built on the architecture of the next-generation Qwen 4 family.
The reported specifications, gathered by the community from the ModelScope listing ahead of release: roughly 125 billion total parameters plus about 51 billion in a separate N-gram embedding component, with only around 6 billion parameters active for any given token. Parameters are the internal numbers a model learns during training, and a token is roughly a word or word-fragment. That gap between 125 billion stored and 6 billion used is the defining feature of a mixture-of-experts design, where the model keeps many specialised sub-networks and only wakes the relevant few for each piece of text. The model is multimodal, meaning it handles images as well as text, and reportedly introduces two architectural changes with the unhelpful names GDN hybrid architecture and Qwen Sparse Attention. Tooling teams including Unsloth have been preparing day-zero support, and both a standard and an FP8 version are expected, FP8 being a compressed number format that roughly halves memory use.
What’s actually going on here: the mixture-of-experts trick is why a model this large is worth anyone’s attention outside a data centre. Running a dense 125-billion-parameter model needs serious hardware; running one that only activates 6 billion at a time is far closer to what a well-equipped desktop can handle, while keeping the breadth of knowledge that comes from the full size. That is the same bet nearly every lab is now making. Releasing a preview of an unreleased architecture as open weights is also a deliberate strategy rather than generosity: it gets the open-source tooling ecosystem, the people who write the software that makes models actually runnable, working on Qwen 4 support before Qwen 4 exists.
What this means for you: if you use AI through a chat window, this changes nothing today, though it feeds into the general trend of the free options getting better. If you run models on your own machine, this is the release to watch this week, with the honest caveat that all the specifications above come from a teaser page and community analysis rather than a published model card, and that day-one support in tools like Ollama and LM Studio usually takes a few days to stabilise. Wait for the first independent benchmark results before rearranging your setup around it.
Sources
Source: https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next
Stanford's Jobs Data Says the Gap for Young Workers Keeps Widening
The August update to the Canaries in the Coal Mine paper puts employment for 22 to 25 year olds in AI-exposed jobs 19 percent below their peers, up from 13 percent. Older workers show no such gap.
