Qwen3.8-Flash-Next
Points and comments are a snapshot, not live.
Qwen3.8-Flash-Next previews Qwen4 architecture with GDN+QSA attention and N-gram embeddings.
Qwen3.8-Flash-Next is a 125B-parameter MoE model (6B activated, plus 51B N-gram embeddings) from Alibaba. It introduces Gated DeltaNet + Qwen Sparse Attention for efficient long-context processing, Gated Residual for cross-layer information flow, and N-gram Embedding for expanded capacity at low compute cost. Trained with Muon optimizer at roughly 1/9 the cost of Qwen3.7-Plus, it outperforms that model on coding and agentic benchmarks. Supports 262K native context, extensible to 1M tokens.
What commenters are saying
Commenters found the benchmarks impressive for the model's size, noting its clean sweep over Qwen3.7-Plus and strong performance against DeepSeek V4 Flash. Several users highlighted relevance for local inference on AMD, Apple Silicon, and consumer hardware. Some pushed back on comparisons with closed models, questioning whether world knowledge gaps and token efficiency offset the cost advantage. A few objected to the frequency of Qwen announcements, but many felt this release justified coverage as a preview of Qwen4's architecture.