Tech
Alibaba ships Qwen3.8-Flash-Next as first Qwen 4 preview at 125B total, 6B active, one-ninth the training cost
MarkTechPost · Bloomberg · 2026-08-26 · Source ↗
Alibaba ships Qwen3.8-Flash-Next as first Qwen 4 preview at 125B total, 6B active, one-ninth the training cost
MarkTechPost · Bloomberg · 2026-08-26 · Source ↗
The Qwen team pushed weights and the technical report to Hugging Face and ModelScope at 15:00 UTC, 30 hours after the Tuesday briefing that first named the model. The architecture is the news. Qwen3.8-Flash-Next is the first shipping instance of the Qwen 4 architecture family, and the design choices in the preview are the choices Alibaba is committing to across the flagship line coming later in the fall.
The parameter budget and the routing math
The backbone runs 125 billion transformer parameters with 512 experts arranged as 10 routed and one shared per layer. Six billion parameters activate per token, giving a sparsity ratio of roughly 4.8 percent. Beyond the transformer sits a 51-billion-parameter N-gram embedding table Alibaba positions as retrieval-native rather than as a bolted-on RAG layer, plus a 4-billion-parameter multi-token-prediction head that lets the model emit multiple tokens per forward pass. The attention layout is hybrid: Gated DeltaNet on three of every four layers combined with Qwen Sparse Attention on the fourth, for a total of 48 layers. Native context is 262,144 tokens and extends to one million with YaRN scaling. The FP8 checkpoint clocks in at 172.78 GiB; BF16 at 335.28 GiB.
"Only 6B parameters activate per token." — Alibaba Qwen team announcement
The benchmark card
The published scores: DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5, CoWorkBench at 73.9, AndroidWorld at 84.5 and MathVision at 95.7 with the code interpreter enabled. Long-context inference speedups run 7.6 times faster on prefill and 4.9 times faster on decode at one-million-token input, which is the number Alibaba is arguing matters for any agent workflow that runs beyond a single turn.
Training cost and the pricing signal
Alibaba disclosed training cost at "roughly one-ninth that of Qwen3.7-Plus." The message is aimed at Nvidia's earnings-day audience: frontier Chinese labs are compressing training cost per capability point at a rate the western frontier labs have not published. The Bloomberg wire wrapped Alibaba shares climbing two percent on the release. The license is qwen-community-1.0, not Apache-2.0, restricting some commercial paths that were open under earlier Qwen 3.x models.
"Training cost roughly 1/9 that of Qwen3.7-Plus." — Alibaba Qwen technical report, Qwen3.8-Flash-Next