Qwen3.8 Flash-Next

Qwen3.8-Flash-Next is Alibaba's experimental preview of the design that Qwen4 will build on, released on August 26, 2026. Its backbone has 125 billion parameters with only 6 billion active per token, and alongside it sit a 51 billion parameter N-gram embedding table, which stores bigram and trigram representations, and a 4 billion parameter multi-token prediction head, so the full checkpoint is about 180 billion parameters on disk. The Qwen team reports it beating Qwen3.7-Plus at roughly one-ninth the training cost. The design combines Gated DeltaNet linear attention with Qwen Sparse Attention, which picks out relevant context in micro-blocks through a lightweight indexer, plus a four-branch gated residual and a 512-expert mixture-of-experts layer that routes 10 experts and one shared expert per token. It accepts text, images, and video, runs in thinking mode by default with the option to turn it off, and supports a 262,144-token native context that extends toward one million tokens with YaRN. Reported scores include 91.7 on GPQA Diamond, 62.5 on SWE-bench Pro, 91.9 on LiveCodeBench v6, and 73.9 on CoWorkBench. The weights are published under the Qwen Community License 1.0 rather than Apache 2.0. It suits coding agents, office and document tasks, and long-context workloads where a very small active-parameter footprint keeps per-token compute light, even though the full checkpoint still needs a multi-GPU node to host.