Alibaba's Qwen team published open weights for Qwen3.8-Flash-Next, a model that handles text, images and video and previews the architecture planned for Qwen4. It activates only a small share of its parameters per request and natively handles very long inputs, with an option to extend further. Running it requires either self-hosted serving software or the paid Qwen Cloud service, and the hosted version is the one offering the full million-token window by default.
What changed
Earlier Qwen releases paired Gated DeltaNet with standard gated attention and did not use n-gram embeddings.
What it unlocks
Self-hosting an open-weight model that handles text, images and video with a very long context and low active parameter count.
- 125B total, 6B activated
- 262,144 token native context
- 1M tokens with RoPE scaling
- 62.5% on SWE-bench Pro
What you need to act on it
- GPU serving infrastructure with vLLM, SGLang or TokenSpeed, or a Qwen Cloud API key
- RoPE scaling configuration for contexts beyond 262,144 tokens
- huggingface.co2026-08-26