Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is Alibaba’s public preview of the architecture that is to carry the next generation of its models. The weights went up on Hugging Face on 24 August 2026 and the hosted version, sold as Qwen3.8-Flash, appeared two days later. The published index totals 180.0 billion parameters, and the model card explains how they split: 125 billion in the language model with 6 billion active per token, plus 51 billion of n-gram embeddings and 4 billion for multi-token prediction. The mixture of experts holds 512 experts, of which ten routed and one shared are used at a time. Attention is hybrid — Gated DeltaNet paired with Qwen Sparse Attention, which selects whole micro-blocks instead of individual tokens, and it is that choice, Alibaba says, that cuts latency on long inputs. The model is multimodal by construction: a causal language model with a vision encoder, taking text, images and video. Native context is 262,144 tokens and can be stretched to a million with YaRN scaling, which the maker warns degrades quality on short inputs; the hosted Qwen3.8-Flash runs at a million by default and costs 0.15 dollars per million input tokens and 0.47 per million output. The licence deserves attention, because the model is widely described as open. Qwen Community License 1.0 permits commercial use, hosting and fine-tuning, but adds two conditions: a product with more than 100 million monthly users or 20 million dollars of monthly revenue must display the model name in its interface, and anyone running a model-as-a-service or an AI coding or office assistant must obtain a separate licence from Qwen first. Internal company use is exempt. This is open weights, not an open licence in the Apache or MIT sense.

▸Videos

No videos yet.

▸Reviews