Alibaba's next open Qwen model nears release as Qwen expands into AI glasses

Code changes landing in public GitHub repositories point to Alibaba's next open Qwen release arriving soon. The previous step was Qwen3.8-Flash-Next, a 125B-parameter open-weight preview of the Qwen 4 architecture.

Qwen 4 is reported to be in active pre-training with several design choices. These include lean-activation mixture-of-experts, dedicated n-gram layers and 1-million-token native context windows. Those choices aim to cut inference cost while extending long-context reasoning, where open models have trailed closed frontier systems.

Alibaba is also moving Qwen into hardware, with two AI glasses models and clip-on AI earphones unveiled. Ecosystem projects such as FLock.io's THIS/THAT edge decision model, built on Qwen3.5-2B, show Qwen weights increasingly serving as a base for other builders.

The open-weight field is crowded this week, with Mistral Large 4 claiming the top open spot and DeepSeek's V4.1-Flash landing in llama.cpp. Observers expect Qwen's next release to be the obvious benchmark comparison for Mistral's claim.