Preview of Qwen4 Model Released via Alibaba Token Plan
Alibaba has initiated a limited preview of its next-generation large language model, Qwen4, through its Alibaba Token Plan. According to reports on Reddit, the preview is currently live across several platforms, including Qwen Office, Qoder, the Qwen App, and QwenStudio. The company is offering the Qwen4-Max Preview at a discount of up to 90% for a limited time.
The company indicates that the new model iteration is designed to provide deeper knowledge, a more comprehensive understanding of global contexts, and enhanced capabilities in coding and agentic tasks. Users can access the preview via the official chat interface at chat.qwen.ai.
Delays and Speculation Regarding Model Capabilities
While the preview is active, the full release of Qwen4 has faced delays beyond its originally scheduled timeline. Official sources have confirmed that the model is “coming soon,” but they have not provided a specific reason for the postponement.
Unverified discussions among the AI community on Reddit have led to various speculations regarding the delay. Some users suggest the timeline was extended to allow for final adjustments based on feedback from an ambassador program. However, other community members have speculated that the delay may be related to the removal of certain technical capabilities. Specifically, some users have alleged that the model’s ability to perform decompiling and reverse engineering tasks has been restricted, potentially to comply with export licensing requirements. These claims regarding “nerfed” capabilities or the removal of specific training data remain unconfirmed by Alibaba.
Technical Observations and Community Feedback
Early testing by users has yielded mixed technical observations. Some users reported encountering a context window limited to 131k tokens, though it remains unclear if this is a limitation of the model itself or a result of specific hardware configurations.
The local AI community continues to monitor the release closely, with significant interest in the availability of quantized weights, such as GGUF formats, and specific parameter sizes like a potential 27b version. Despite the delays and the speculative nature of the model’s current constraints, the Qwen series maintains a significant presence in the local AI development sector.
