Qwen3.8-Flash-Next: the open MoE with 6B active params previewing Qwen 4
Alibaba has open-sourced Qwen3.8-Flash-Next, a multimodal MoE model with 125B parameters and only 6B active per token that serves as an official preview of the architecture behind Qwen 4.
What is Qwen3.8-Flash-Next
Released on August 26, 2026, Qwen3.8-Flash-Next plays the same role that Qwen3-Next played for the 3.5 series: opening up architectural changes early so the community can examine them before the full Qwen 4 family is built on top. The model adds 51B parameters in n-gram embeddings and 4B in an MTP module, with 48 layers and 512 experts, activating 10 routed plus 1 shared per token.
The architecture of Qwen3.8-Flash-Next
The changes concentrate on four fronts. For attention, it combines Gated DeltaNet (GDN) with Qwen Sparse Attention (QSA): instead of selecting individual tokens, QSA indexes context at micro-block granularity with a compressed lightweight indexer, substantially cutting long-context attention cost. For residuals, Gated Residual widens the residual stream into 4 branches with dynamic read and write gates. N-gram embedding looks up a table of 20 million bigrams and trigrams to scale capacity with very little extra compute, and the table can be offloaded to host memory. For optimization, the training recipe combines the Muon and AdamW optimizers with refitted scaling laws, removing batch-size warmup entirely.
Native context is 262,144 tokens, extensible up to 1 million, and the model is multimodal: it ships a vision encoder for image input.
Results against much larger models
With only 6B active parameters, Qwen3.8-Flash-Next outperforms models that are far more expensive to serve. It scores 58.7 on DeepSWE 1.1, versus 54.4 for DeepSeek-V4-Flash-0731 and 16.5 for Qwen3.7-Plus; 62.5 on SWE-bench Pro, beating Qwen3.7-Plus (55.8) and Claude Opus 4.6 (53.4); and 81.0 on SWE-bench Multilingual. It also shines on multimodal agent tasks, with AndroidWorld 84.5 and Vision2Web 64.0.
Efficiency and training cost
Alibaba reports that training Qwen3.8-Flash-Next cost about one-ninth of Qwen3.7-Plus, while delivering stronger coding and office-task capabilities. The production version, Qwen3.8-Flash, adds a 1M context window by default and built-in tools.
Availability and ecosystem
Weights are published on Hugging Face and ModelScope under the qwen-community-1.0 license. It can be served with Transformers, llama.cpp, Unsloth, SGLang, vLLM, and TokenSpeed, and it is already live in QwenWork (Standard mode), on the QwenCloud API compatible with OpenAI and Anthropic, and in the Qwen Code terminal agent.
Conclusion
Qwen3.8-Flash-Next confirms Alibaba's bet on far more efficient open models and gives us a clear look at what to expect from Qwen 4. If you missed the Qwen3.8-27B release, that story rounds out the series picture.
