Unconfirmed: A post on X claims details about Alibaba's upcoming Qwen4 model tiers and a potential release window, but no publisher has confirmed the report.
@pratikg on X1 wrote on 9 November 2026 that Alibaba had named four Qwen4 tiers—Max, Plus, Flash and 27B—during a presentation, with only the 27B described as open weights. The post also cited leaked kernel code indicating a dense architecture with sparse attention and a built-in drafter, and noted that some developers were already experimenting with Qwen3.6 27B fine-tunes. Hardware requirements were discussed, with 4-bit quantization of Qwen4 27B estimated at ~16 GB and 8-bit at ~29 GB, implying an RTX 3090 (24 GB) could run the 4-bit version but not the 8-bit version.
The post further claimed that GameFeel scores for Qwen4 27B were achieved using early access to an API rather than open weights, and that the model narrowly outperformed DeepSeek V4.1 Flash 1017 on a single benchmark. It also disputed claims of 5–10 trillion parameters for Qwen4, attributing such figures to future Qwen 4.5 or Qwen 5 releases. A rumored release time of 3 PM Beijing time on 11 November 2026 (11 PM PT on 10 November) was described as unconfirmed, with the author stating they would only believe it upon seeing the model on Hugging Face.
The post also speculated on day-zero MLX performance for Qwen4 27B on Apple Silicon and discussed potential performance implications of sparse attention beyond 2,000 tokens, noting that KV cache size may not shrink despite sparse attention. The author invited suggestions for first tests once the model is available.
No independent confirmation of these claims has been published, and Alibaba has not publicly commented on the details described in the post.
What would confirm or kill the rumour: An official announcement or model release from Alibaba Qwen, or a verified posting of open weights for Qwen4 27B on a platform such as Hugging Face. The absence of such confirmation keeps the report unconfirmed.