StudioQwen Releases Experimental Preview Model Qwen3.8-Flash-Next to Pave the Way for Qwen4
Original title: Qwen/Qwen3.8-Flash-Next (image-text-to-text)
Qwen has released an experimental architecture preview model, Qwen3.8-Flash-Next, designed to support the upcoming Qwen4 architecture. This image-to-text and text-to-text model features a total of 125 billion parameters (with 6B activated), including 51B n-gram embeddings and 4B MTP. It introduces Qwen Sparse Attention (QSA), Gated Residual connections, and N-gram Embeddings, supporting a native context window of 262,144 tokens, scalable up to 1 million. Official claims state that QSA operates at the micro-block level to significantly reduce latency in long-context scenarios. Weights are compatible with Transformers, vLLM, and SGLang, while the official version based on this architecture, Qwen3.8-Flash, is available via Qwen Cloud.
The original article provides detailed architectural specifications such as QSA, Gated Residual, and N-gram Embedding, allowing readers to understand the technical direction of Qwen's pre-Qwen4 architecture.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.