Quick answer
Qwen released Qwen3.8-Flash-Next on August 26, 2026 as an open-weight multimodal mixture-of-experts model and an experimental preview of the architecture planned for Qwen4. Qwen reports a 125B-parameter main model with 6B parameters activated per token and a native 262,144-token context window. Availability in Chat AI has not been verified.
What Qwen released
Qwen3.8-Flash-Next is a post-trained multimodal mixture-of-experts model with public weights and configuration files. Qwen calls it an experimental preview of the architecture intended to underpin Qwen4, rather than the complete Qwen4 model family. The official release repository and model card were published on August 26, 2026.
Sources: Qwen Team, Qwen on Hugging Face
The architecture changes focus on long-context efficiency
The release combines Gated DeltaNet with Qwen Sparse Attention, which selects context in small blocks instead of processing every token with full attention. Qwen also introduces gated residual streams, short n-gram embeddings, and a training recipe that assigns Muon and AdamW optimizers to different weight categories. These are architectural changes intended to reduce long-context and training costs while preserving model capacity and stability; real-world gains will depend on the workload and serving stack.
Sources: Qwen, Qwen Team, Qwen on Hugging Face
Model size and context window
Qwen describes the main language model as having 125B parameters with 6B activated per token, plus 51B n-gram embedding parameters and a 4B multi-token-prediction component. The model card lists a native context length of 262,144 tokens and says it can be extended to 1,000,000 tokens. Those figures describe the official architecture and supported configuration, not a guarantee of equal speed or quality on every deployment.
Sources: Qwen Team, Qwen on Hugging Face
How the model can be accessed
The official weights are available through Hugging Face and ModelScope, while Qwen lists Transformers, llama.cpp, SGLang, vLLM, TokenSpeed, and other tools for local or hosted inference. The release uses the Qwen Community License rather than Apache 2.0, so teams should review the license before deployment. Qwen also says a production-oriented Qwen3.8-Flash service based on this architecture is available through Qwen Cloud with additional built-in features.
Sources: Qwen Team, Qwen on Hugging Face
Treat the published benchmarks as first-party results
Qwen reports results across coding, agent, general reasoning, and vision-language evaluations, and provides some evaluation settings alongside the tables. These are useful release evidence, but they remain publisher-run results. Compare the model on representative prompts, tools, files, images, latency, memory use, and total cost before moving a production workflow.
Sources: Qwen, Qwen on Hugging Face
Qwen3.8-Flash-Next is not yet verified in Chat AI
Qwen3.8-Flash-Next does not appear in Chat AI's public model directory at publication time, so this article does not claim that the model is available in the app. Check the current model directory and in-app selector before assuming access; availability can vary by plan, platform, region, provider access, and app version.
Sources: Chat AI
Frequently asked questions
What readers usually ask
Is Qwen3.8-Flash-Next the same as Qwen4?
No. Qwen describes it as an experimental preview of the architecture planned for Qwen4, not as the full Qwen4 model family.
Is Qwen3.8-Flash-Next open source?
The model weights are publicly available, but they use the Qwen Community License rather than Apache 2.0. Review the license terms for your intended use instead of assuming unrestricted open-source rights.
What context length does Qwen3.8-Flash-Next support?
The official model card lists a native context length of 262,144 tokens and says the model can be extended to 1,000,000 tokens. Hardware and serving requirements still depend on the deployment.
Is Qwen3.8-Flash-Next available in Chat AI?
Availability in Chat AI was not verified at publication time. Check the current model directory or in-app selector before assuming access.
Evidence