Qwen released Qwen3.8-Flash-Next, an open-weight model that handles both text and images and previews the architecture planned for its next generation. It uses a sparse design in which only a small share of its parameters run for any given request, which the team says cuts delay on very long inputs. The published figures put it ahead of larger rivals on several coding and agent tests, and using it requires self-hosting or the paid Qwen Cloud service.
What changed
Qwen's prior open-weight models used a Gated DeltaNet and Gated Attention pairing with smaller context and higher long-context latency.
What it unlocks
Running a text-and-image model with a very long working memory on own hardware, using common open serving software.
- 125B parameters, 6B active
- 262,144 tokens native context
- up to 1,000,000 tokens with scaling
- 62.5 on SWE-bench Pro
What you need to act on it
- hardware and skills to self-host, or the Qwen Cloud paid API
Sources