Qwen 3.8 Flash Next is presented as an experimental preview of ideas expected in Qwen 4. It uses a 125-billion-parameter mixture-of-experts designA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. while activating roughly 6 billion parameters for each tokenAn active AI model parameter is a model weight selected to participate in computation for a particular token or inference step., aiming to keep inference compute low despite the model's total size.
The architecture mixes three gated-delta linear-attention blocks with one sparse-attention block. It also assigns about 51 billion parameters to n-gram lookup embeddings that can live in ordinary system memoryAn n-gram lookup embedding stores learned representations for recurring token sequences so an AI model can retrieve them without recomputing every pattern.. That arrangement reduces repeated computation for common token patterns, but it shifts a substantial part of the cost into RAM capacity and bandwidth.
The host's eight-test comparison scored Qwen at 56 out of 80, or 70 percent, against 63 out of 80, or 78.75 percent, for GLM-5.3 Flash. Qwen performed well on math, an agentic pipeline and one 3D task, while GLM produced more polished results in several visual and front-end tests. These are the host's benchmark results rather than a broad independent evaluation.
Running a useful local quantization may require roughly 96 to 128 GB of memoryA model memory footprint is the total working memory required to load and run an AI model in a chosen configuration.. That makes the model compute-efficient but memory-hungry, and it can leave API access cheaper than buying suitable hardware. The reported 262,000-token native contextA context window is the maximum amount of tokenized information an AI model can consider during one processing session. and optional one-million-token YaRN extension are vendor claims that still need practical validation across real workloads.
Watch on YouTube
