Qwen3.8-Flash-Next

Qwen3.8-Flash-Next, released August 26, 2026, is Alibaba's open-weights preview of the architecture that will underpin Qwen4. The headline reads 125B with 6B activated, but the full checkpoint is roughly 180B: a 125B transformer stack, a 51B n-gram embedding table that lives in host RAM via row lookups, and a 4B multi-token-prediction head. It combines Gated DeltaNet linear attention with a sparse-retrieval mechanism (Qwen Sparse Attention), a 512-expert MoE with a gated shared expert, and Muon-optimizer training. Native context is 262,144 tokens, extensible toward 1M. Alibaba reports 62.5 on SWE-bench Pro from its own agent-harness runs and claims training cost roughly one-ninth of Qwen3.7-Plus.

The honest picture: all headline numbers are vendor-supplied, and the model is explicitly experimental — an architecture preview, not a finished Qwen4 product. The checkpoint is heavy (360GB full, ~186GB FP8), so single-consumer-GPU inference is unrealistic despite the 6B-active headline. The Qwen Community License 1.0 is free for most uses but requires a separate agreement to sell model-as-a-service access.