Qwen Opens Flash-Next Weights as a Preview of Its Qwen4 Architecture

Qwen has released Qwen3.8-Flash-Next, an experimental open-weight model that previews the architecture planned for Qwen4. The August 26, 2026 release is covered by LLM Stats, which notes that this is a self-hosting model—not the separately managed Qwen Cloud Flash product.

The model combines text, image, and video input with text output. It has roughly 180 billion stored parameters, but activates 6 billion, alongside a 51B n-gram table and 4B multi-token prediction component. Its context window is 262,144 tokens natively, with support for up to 1 million tokens through YaRN. The Hugging Face card lists the qwen-community-1.0 license.

Why builders should care

Flash-Next targets teams running their own inference stack with Transformers, vLLM, SGLang, or TokenSpeed. Its hybrid design pairs Gated DeltaNet with Qwen Sparse Attention, while the n-gram embedding system is intended to provide scale that may be cheaper to offload than conventional expert parameters.

Qwen’s published—but not independently verified—benchmarks show gains over DeepSeek-V4-Flash-0731 on several coding and agent tests: DeepSWE 1.1 scores 58.7 versus 54.4, and SWE-bench Pro reaches 62.5 versus 56.0. LiveCodeBench v6 is 91.9 versus 90.6, while GPQA Diamond is 91.7 versus 90.8. The result is weaker on NL2Repo-Bench, where it records 48.1 versus 54.2.

There is no hosted price yet, and the weights were not listed with an inference provider at publication. Developers wanting official tools and 1M context by default should look to Qwen Cloud Flash instead; Flash-Next is an infrastructure project and architecture preview, not a drop-in production endpoint.

Source: LLM Stats