Qwen released Qwen3.8-Flash-Next, an open-weight preview of the Qwen4 architecture

Archive entry: this covers an event from August 26, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On August 26, 2026, Qwen released Qwen3.8-Flash-Next, an open-weight multimodal MoE model that it presented as an early preview of the Qwen4 architecture. The Qwen3.8-Flash-Next README on GitHub listed 125B parameters, of which 6B were activated.

What changed

• Size: 125B parameters with 6B activated, plus a 51B n-gram embedding and a 4B MTP module, 48 layers and 512 experts (10 routed plus 1 shared).
• Context: 262,144 tokens, extensible to 1M.
• Architecture: GDN combined with Qwen Sparse Attention (QSA), a Gated Residual and an N-gram Embedding that can be offloaded to host memory. The model was trained with the Muon optimizer.
• The license was qwen-community-1.0, and an FP8 variant was available.

Hosted version and tooling

Qwen3.8-Flash was the official hosted version based on Qwen3.8-Flash-Next, with 1M context and built-in tools. The Alibaba Cloud Model Studio entry for August 26, 2026 listed OpenAI and Anthropic protocols for qwen3.8-flash and use with tools such as Claude Code. Qwen also listed the model on its Hugging Face model page.

What it means for people who build with AI

The open release gave developers a look at Qwen4 design choices, such as the combination of GDN with QSA and an n-gram embedding that can be offloaded to host memory. The hosted version could plug into Claude Code through an Anthropic-compatible API.

Sources

• Qwen3.8-Flash-Next (GitHub README) (Alibaba (Qwen), August 26, 2026, official)
• Qwen3.8-Flash-Next (Alibaba (Qwen), official)
• Alibaba Cloud Model Studio:Model lifecycle and updates (Alibaba Cloud, August 26, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.