Reportedly: Alibaba releases Qwen4-Max, a 2.4-trillion-parameter MoE flagship model with 1M-token context and over 60% lower task cost
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Alibaba's Qwen team released Qwen4-Max, a 2.4-trillion-parameter sparse MoE flagship model with native vision and a 1M-token context window. It is available via the Qwen AI platform through Token Plan or API. The company reports task cost reduction of over 60% compared with Qwen3.8-Max.
China context
- Original name
- 通义千问 Qwen4-Max
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China should verify API availability and pricing in their region; the evidence only confirms access via the Qwen AI platform, which may be China-focused.
- For investors
- Alibaba's reported cost reduction and planned open-sourcing of smaller models could intensify price competition among Chinese model providers, affecting margins for API-based businesses.
Translated from Chinese. Quotes and facts link to the original sources.
Alibaba's Qwen team announced the official release of Qwen4-Max, a new flagship model using a sparse Mixture-of-Experts (MoE) architecture with 2.4 trillion backbone parameters. The model natively supports visual understanding and provides a 1 million token context window. It is now available on the Qwen AI platform, with developers able to access it through Token Plan or API. The architecture includes Gated DeltaNet in three of every four layers to compress historical information into a fixed state, Qwen Sparse Attention in the remaining layer to select relevant context by block, Gated Residual with four parallel branches, and N-gram Embedding in the second layer. Alibaba reports that the model reduces token consumption and task cost by over 60% compared with Qwen3.8-Max, while improving performance on programming, office, and long-horizon tasks. The company also stated that smaller-sized models will be open-sourced later.
Qwen4-Max uses a sparse MoE architecture with 2.4 trillion backbone parameters. It combines Gated DeltaNet layers (three out of four) for fixed-state historical compression, Qwen Sparse Attention for block-level context selection, Gated Residual with four parallel branches, and N-gram Embedding for local token patterns. The model supports a 1M-token context window and native vision. Alibaba reports over 60% lower task cost and improved performance on programming, office, and long-horizon tasks compared with Qwen3.8-Max. Independent benchmarks are not yet available.
Developers using Qwen's API face lower per-task costs and a longer context window, which may reduce the need for external retrieval or summarization pipelines. Alibaba's move to open-source smaller Qwen4 models later could pressure other Chinese labs to match both pricing and weight availability.
The model is available via Qwen AI platform Token Plan or API. The reported 60%+ cost reduction could lower enterprise spending on complex tasks. Native vision and 1M context may enable new applications in document processing and multimodal workflows.
Watch for the release of smaller Qwen4 open-weight models and independent benchmark results. A key verification question is whether the reported cost reduction holds across different task types and API usage patterns.