Qwen opened a 125B model with 6B active, and the quantized file is what decides if you can run it

Qwen released the weights of Qwen3.8-Flash-Next on August 26, and described it as an early preview of the architecture that Qwen4 will be built on. For anyone deciding what to run at home, two numbers separate this from the usual open-weights announcement: the model has 125B parameters in its main body, and activates 6B of them per token.

What has to fit in memory

The main model is 125B parameters, supplemented by an additional 51B of N-gram embeddings, with 6B parameters activated per token. The sparse activation is what makes generation fast on a modest machine. The total is what makes it not fit, and that is the constraint you actually plan around.

Simon Willison, who publishes his own experiments rather than reporting other people's, has been running it on a DGX Spark using Unsloth quantized builds. He names two: a 72.5GB UD-IQ1_S and a 78.9GB UD-Q2_K_XL. Those file sizes are the requirement in its most useful form. This is a machine with roughly 80GB of usable memory, which means a DGX Spark or a high-memory Mac, not a laptop.

One architectural decision points straight at this case. Qwen says the N-gram embedding table can be offloaded to host memory and overlapped with model computation through asynchronous prefetching, so the largest single block of parameters does not have to sit in the accelerator alongside everything else.

The architecture Qwen4 will inherit

Four changes carry the release. Attention is a hybrid: three of every four layers use Gated DeltaNet to compress history into a fixed-size state, and the remaining layer uses global attention, now with Qwen Sparse Attention selecting important context at micro-block rather than token granularity. Gated Residual widens the residual stream into four branches under a dynamic gate. N-gram Embedding scales capacity through a lookup rather than through computation. The Muon optimizer replaces the previous setup, with the scaling law refitted for the new architecture.

Qwen says the result costs about one ninth as much to train as Qwen3.7-Plus, which carries 397B parameters with 17B active. Context is 262,144 tokens natively and extends to 1,000,000 with YaRN. The open weights are on Hugging Face and ModelScope, and a hosted production version is served on QwenCloud.

Where it lands on the benchmarks

Qwen's own table puts it at 62.5 on SWE-bench Pro against 55.8 for Qwen3.7-Plus and 53.4 for Claude Opus 4.6, at 81.0 on SWE-bench Multilingual, and at 58.7 on DeepSWE 1.1. On the general set it reports 91.7 on GPQA Diamond and 91.9 on LiveCodeBench v6. On the multimodal side, AndroidWorld is 84.5.

These are the vendor's own measurements, run with the vendor's harness choices, and they should be read as such until someone independent repeats them. What they do establish is the shape of the claim: a model activating 6B parameters is being placed against models activating 17B.