Qwen3.8-Flash-Next - The open-weight preview of Qwen4

Hi everyone!

Qwen has done this once before. Qwen3-Next gave everyone an early look at the architecture that later showed up in Qwen3.5.

Qwen3.8-Flash-Next is doing the same for Qwen4.

It’s a 125B model with only 6B active parameters, and a lot of the new design is about getting more capability without dragging compute up with it. QSA makes long-context retrieval cheaper, Gated Residual gives information more paths through the network, and the new N-gram memory adds capacity with very little per-token compute.

Qwen keeps putting these architecture previews out as real models people can actually run, well before the next generation arrives.

And the weights are already up :) (with a license worth reading)

Congrats on the launch! love that these previews are actual downloadable models.

the active param count is spot on — 125B total is what kills it for local though. any chance of a ~35-40B/A3-6B version on this architecture? 35B-A3B was the local goat and it has no successor yet 🙏

I think this is the best local model which can run with 128GB memory.

the gated residual + n-gram memory combo is the part i'd want to poke at. changing how information flows through the network usually means old LoRA adapters and fine-tunes built for Qwen3 don't transfer cleanly to the new architecture. is that the tradeoff here, or did you design it so existing Qwen3 tooling and adapters still mostly work on top of this