Qwen3.8-Flash-Next
Alibaba · Released Aug 2026
Conditional. Qwen3.8-Flash-Next is Alibaba's open-weight experimental preview of the architecture it says will underpin Qwen4: 125B parameters with 6B activated, 262,144 native context extensible to 1,000,000, and image and video input. The deciding caveats are its preview status, the Qwen Community License 1.0 (not Apache) with a separate-licence clause for Model as a Service businesses, and no stated price. Suits researchers and teams that can self-host a large model and want to test Qwen4's direction.
Is it right for you?
Good for
- Evaluating the architecture Alibaba says will underpin Qwen4; the model card calls it an experimental preview of that architecture
- Self-hosted multimodal work; the card describes a causal language model with a vision encoder, gives image and video examples, and recommends SGLang, vLLM or TokenSpeed for serving
- Long-context experiments; the card gives 262,144 tokens natively and says it is extensible to 1,000,000
- Agentic coding trials; Qwen's own table lists SWE-bench Pro 62.5 against 61.7 for Qwen3.8-27B and 56.0 for DeepSeek-V4-Flash-0731 (vendor numbers, untested by us)
Not good for
- Production dependence on stable behaviour; the card itself labels the model an experimental preview
- Teams that need Apache or MIT terms; the licence is the Qwen Community License 1.0, which requires showing the model name above 100,000,000 monthly active users or US$ 20,000,000 monthly revenue and requires a separate licence for Model as a Service or AI Work Assistant businesses
- Budgeting a hosted API; the card names Qwen Cloud as the official API service, and the pages we read state no price
- Decisions that need independent benchmark confirmation; every score comes from Qwen's own model card and we have not tested the model
How it performs by task
Coding agents
SWE-bench Pro 62.5 on the model card, ahead of Qwen3.8-27B (61.7) in Qwen's own table; vendor claim, untested by us
Reasoning
GPQA Diamond 91.7 on the model card against 89.2 for Qwen3.8-27B and 90.8 for DeepSeek-V4-Flash-0731; vendor claim, untested by us
Multimodal understanding
the card lists image and video input with usage examples; we read no vision score we could check twice, so this rating rests on the stated capability only
Pricing
Input
Not stated
Output
Not stated
Context
Open weights; Qwen Community License 1.0
Benchmarks
No verdict changes yet
The clock starts day one. Changes land here as our verdict evolves.
Sources
Verification log
- Status· No changes
Automated agent
- Pricing· No changes
Automated agent