Qwen3.8-Flash

Qwen3.8-Flash is the hosted build of Qwen3.8-Flash-Next, the open-weight model Alibaba published on August 26, 2026 as an explicit preview of the architecture that will underpin Qwen4. Keep the two straight: Flash-Next is the downloadable preview under a qwen-community-1.0 license with a 262,144 token native window and no list price, while Qwen3.8-Flash on QwenCloud is the managed product at $0.16 per million input tokens and $0.47 output with a 1 million token default context and official tool support. The architecture is the interesting part. It is 125 billion parameters with only 6 billion activated per token, paired with a 51 billion parameter n-gram embedding table indexing 20 million bigrams and trigrams, and a 4 billion parameter multi-token prediction layer. The n-gram table is the unusual lever: those parameters can sit in ordinary system RAM instead of GPU memory, which makes them far cheaper to hold than MoE experts. Alibaba reports better results than the 397 billion parameter Qwen3.7-Plus at roughly one ninth the training cost. On the self-reported table against DeepSeek-V4-Flash-0731 it leads DeepSWE 1.1 (58.7 to 54.4), SWE-bench Pro (62.5 to 56.0), Toolathlon Verified (73.5 to 70.3), GPQA Diamond (91.7 to 90.8), and LiveCodeBench v6 (91.9 to 90.6). NL2Repo-Bench is the honest miss at 48.1 against 54.2. Vision numbers are strong on paper (AndroidWorld 84.5, RealWorldQA 88.5, MathVision 90.6) but OSWorld 2.0 sits at 19.4, which is still a coin flip on full task success. Every figure here is vendor-reported.