Alibaba releases Qwen3.8 preview: dynamic MoE and Sparse-Tiled Attention, open-source version imminent

On September 12, 2026, Alibaba Tongyi Lab released a Qwen3.8 early-access preview and confirmed that the full release will be fully open-sourced. The version is a systematic rebuild across base architecture, inference mechanism, and multimodal fusion, aimed at pushing the capability ceiling of open-source large models into the next efficiency tier. At the base layer, Qwen3.8 uses an improved Mixture-of-Experts architecture with a new Dynamic Router v4. Traditional MoE routers typically rely on static thresholds or simple attention scores, which leads to uneven load. Qwen3.8's reinforcement-learning-trained router assigns compute to specific expert networks in real time based on the complexity of the input token. Simple conversational traffic activates only a tiny fraction of parameters and runs at very low latency, while complex code generation or multi-step logical reasoning automatically activates more high-tier experts. Qwen3.8 also introduces Sparse-Tiled Attention, a variant of Flash Attention-3, that significantly improves memory access efficiency for long context windows. Combined with KV-cache quantization, the model cuts VRAM usage by nearly 40% on 128K-and-above contexts. This engineering optimization lets trillion-parameter models run efficiently on consumer-grade GPU clusters, materially lowering the compute bar for enterprise deployment. On the inference side, Qwen3.8 introduces a CoT++ self-verification loop that automatically validates and rolls back intermediate steps on long-chain tasks, reducing multi-step failure rates. On the multimodal side, a UMT (Unified Multimodal Transformer) embedding maps text, image, audio, and video into a shared vector space, supporting native cross-modal retrieval and generation and avoiding the feature fragmentation of bolted-on multimodal stacks. The Qwen3.8 preview is available for trial on the Qianwen AI platform; the full release will open-source both the base model and the fine-tuning weights.