Qwen3.8 Flash Next

Qwen3.8-Flash-Next is an experimental preview of the architecture that will underpin Qwen4, built around a fundamental rethinking of how the core components of modern large language models interact at scale.

Highlights

• Hybrid Attention with QSA: Gated DeltaNet is paired with Qwen Sparse Attention (QSA), which operates at the micro-block level rather than selecting individual tokens. This significantly reduces long-context latency for agentic workloads.
• Efficient Sparse Scaling: the language model has 125B parameters with 6B activated, plus a 51B n-gram embedding and 4B MTP. N-gram embeddings provide a compute-efficient scaling axis that is especially amenable to offloading.
• Gated Residual: widened residual streams use data-dependent read gates and per-branch scalar write gates for greater expressiveness while preserving training stability.
• Vision and Agentic Capabilities: native image understanding complements strong performance across coding, professional work, long-horizon agents, and multimodal tool use.
• Flexible Thinking Control: thinking is enabled by default, with xhigh, medium, and low reasoning-effort levels. Reasoning from previous messages is preserved by default for continuity in multi-turn agentic workflows.
• Long Context: natively supports up to 262,144 tokens.