ModelScope has posted a teaser page for Qwen3.8-Flash-Next, with a countdown pointing to 23:00 (UTC+8) on August 26. Both the standard and FP8 versions will open for download at the same time. On Hugging Face, the description for the same-named repository is a single line: "A Preview of the Qwen4 Architecture."
The parameter configuration leaked from the teaser page: 125B main parameters, an additional 51B N-gram embedding, and 6B activated per token. The model is positioned as a multimodal MoE. Officially, it is built on the next-generation Qwen4 architecture, open-sourced ahead of schedule to help the community prepare for the Qwen4 series.
Activation Rate Pushed Below 5%
Only 6B out of 125B gets activated, putting the activation rate at roughly under 5%. For reference, the sparse MoE models common across the industry over the past two years mostly land in the 5%-to-10% range — Flash-Next pushes the sparsity even lower.
This is a configuration clearly aimed at inference cost. MoE memory footprint is determined by total parameters, while compute load is determined by activated parameters — the 125B of weights still has to fit in VRAM, but each forward pass only routes through a 6B path. On the deployment side, that means: the GPU capacity threshold does not drop, but the compute and latency threshold drops sharply. Paired with the FP8 version, weight footprint can be cut roughly in half again. The "Flash" in the name most likely refers to this.
One unusual item in the parameter table is the 51B N-gram embedding, listed separately from the main parameters. Embedding-type structures are typically lookup-table-based and do not participate in layer-by-layer matrix operations; if Flash-Next follows the same approach, the 51B looks more like spending VRAM to buy performance than added compute — though exactly how it is implemented will not be confirmed until tonight's model card and config files are released.
Why "3.8" Carries the Qwen4 Architecture
The naming looks awkward, but the logic tracks. This is not the first time Qwen has released a Next preview near the end of a 3.x series — the playbook is the same: run a mid-size model through the next generation's routing method, attention variant, and training stack, open-source it, let the community get the toolchain ready first, and only then roll out the formal major version.
The payoff is concrete. Whether an open-source model can be picked up directly by frameworks like vLLM, SGLang, and llama.cpp on release day largely determines its adoption curve over the first two weeks. Once an architecture introduces new operators or a new routing method, adaptation can drag on for weeks. Unsloth has already announced it is working on day-zero support — following exactly this path, moving the time cost of architecture adaptation from Qwen4's release day to today.
Before the repository has even opened, 1,299 people on Hugging Face have already clicked to be notified — not a low number for an empty repo that does not even have a model card yet.
What to Watch After Tonight
The teaser page only goes as far as parameter count — several key metrics are still undisclosed: context length, which modalities the multimodal coverage extends to, whether the license continues under Apache 2.0, and, most importantly, benchmark scores.
Over the past year, Alibaba's open-source playbook has treated weights as an entry point into its ecosystem, releasing even Max-tier model weights. Flash-Next pushes the timeline further forward still: releasing the architecture before the model itself is mature. For teams in China doing inference deployment and fine-tuning, the 48 hours after 23:00 tonight are going to be busy.
Sources: ModelScope model page, Hugging Face repository, CocoLoop, Decrypt, Hacker News community discussion; parameter configuration and release time are as shown on the model page, and the activation rate is a rough estimate based on public parameters.