Qwen3.8-Flash-Next
Qwen3.8-Flash-Next is Qwen's experimental open-weight preview of the architecture intended to underpin Qwen4.
Provider
Alibaba Qwen
Model family
Alibaba Qwen
Open-weight multimodal MoE LLM
Cost tier
Flash Next
Status
Current
Release Aug 26, 2026
Why teams choose it
Qwen describes this checkpoint as an experimental preview of the Qwen4 architecture
not a drop-in replacement for hosted Qwen3.8-Max.
The hosted Qwen3.8-Flash service is documented as a production SKU based on this checkpoint
with 1M context by default and built-in tools.
Native context is 262,144 tokens
with documented extension toward 1M under supported serving configurations.
Tradeoffs to know
- This is an experimental weights preview; production features such as default 1M context and official tools live on hosted Qwen3.8-Flash.
- The 51B n-gram table and MoE layout change memory and serving behavior versus dense Qwen3.8-27B.
- License terms are the Qwen Community License rather than Apache 2.0; review commercial-service obligations on the model card.
Technical specs
- Inputs
- text, image, video
- Outputs
- text
- Capabilities
- coding, agentic workflows, multimodal reasoning, long-context, open-weight serving, configurable reasoning
- License
- Qwen Community License 1.0
- Source
- https://huggingface.co/Qwen/Qwen3.8-Flash-Next
- Native context
- 262,144
- Total parameters
- 125B
- Activated parameters
- 6B
Benchmarks
No independent or vendor benchmark scores in this catalog entry.
Alibaba Qwen family lineup
Current models
Related models
Qwen3.8-Flash-Next FAQ
What is Qwen3.8-Flash-Next?
Qwen3.8-Flash-Next is Qwen's experimental open-weight preview of the architecture intended to underpin Qwen4. The official Hugging Face card documents a 125B MoE with 6B activated parameters plus 51B n-gram embeddings and a 4B MTP head, native 262,144-token context extensible to 1M, and text, image, and video input.
When does Qwen3.8-Flash-Next fit best?
Local or private multimodal agents
What should teams watch out for with Qwen3.8-Flash-Next?
This is an experimental weights preview; production features such as default 1M context and official tools live on hosted Qwen3.8-Flash.
Explore next
Models, tools, and comparisons that connect to this reference.