News AI
Qwen4-Max Alibaba Debuts as 2.4 Trillion Parameter MoE Model with 1M Token Context
Alibaba made Qwen4-Max generally available on November 16, 2026, introducing a sparse MoE model with 2.4 trillion parameters and a 1 million token context window.
Alibaba released Qwen4-Max on November 16, 2026, as the first model in its Qwen4 series and the new flagship successor to Qwen3.8-Max. The model became available through general access on the QwenCloud API after a limited Token Plan preview that began November 11.
Qwen4-Max Alibaba Technical Specifications
The architecture uses a proprietary sparse Mixture-of-Experts design with a 2.4 trillion parameter backbone. It activates 71 billion parameters per token, which represents a 25 percent reduction compared with the active parameters of Qwen3.8-Max at equivalent scale. An additional 328 billion N-gram embedding parameters support the model. Key upgrades in the Qwen4 architecture include hybrid attention combining Gated DeltaNet with Qwen Sparse Attention, gated residuals, and N-gram embeddings. The system accepts native multimodal inputs in text, images, and video while producing text output.

Context length reaches 1 million tokens. Maximum output extends to approximately 393,000 tokens when thinking and answer tokens are combined, though limits are lower without the thinking mode enabled. These specifications target long-running agentic workloads and extended professional tasks that require retention of large context across multiple steps.
Alibaba positions the model for coding, visual reasoning, scientific discovery, and enterprise automation. The efficiency improvements allow the same or better output quality while consuming fewer tokens on many internal benchmarks, which directly reduces per-task API costs for users.
Access Options and Updated Pricing
Qwen4-Max remains hosted only, with no open weights released for the Max variant. Users can access it via the QwenCloud API using the model identifier qwen4-max, as well as through Qwen Studio, QwenWork, Qoder, and the Token Plan. The service supports OpenAI- and Anthropic-style API calls, including function calling and structured outputs.
International list pricing stands at 1.50 dollars per million input tokens and 4.50 dollars per million output tokens, a reduction of roughly 25 percent from Qwen3.8-Max rates. Implicit cache input tokens are priced at 0.075 dollars per million, which is about 70 percent lower than the prior rate. These changes make high-volume or long-context usage more economical for developers and organizations running repeated or extended queries.
Smaller open-weight Qwen4 models are expected soon, though Alibaba has not announced specific release dates or parameter sizes. The current release focuses on enterprise and developer access through paid API tiers rather than local deployment.
Reported Performance and Efficiency Gains
Alibaba states that Qwen4-Max outperforms its predecessor on the internal Work360 benchmark covering 721 specialized fields. The model achieves the results while using approximately 35 percent fewer tokens at high reasoning effort, which translates to average task cost reductions of 60 to 62 percent. Some individual tests show token savings approaching 50 percent. The model records strong scores on agentic, coding, and visual reasoning tasks, topping four of the 16 benchmarks shown in company charts, including a score of 21.3 on HLAB-Vals-1.1.
Independent third-party evaluations place the model competitively among current large language models, though it trails certain US and Chinese rivals on specific leaderboards. The efficiency claims stem from architectural changes that optimize attention mechanisms and residual connections for long-context and multi-step reasoning workloads. Organizations evaluating the model can test these metrics directly through the pay-as-you-go API pricing now in effect.
Frequently asked questions
When did Alibaba release Qwen4-Max?
Alibaba made Qwen4-Max generally available on November 16, 2026, following a limited preview that started November 11.
What is the parameter count of Qwen4-Max?
The model has a 2.4 trillion parameter backbone and activates 71 billion parameters per token during inference.
How large is the context window in Qwen4-Max?
Qwen4-Max supports a native context window of 1 million tokens with maximum output reaching up to 393,000 tokens when thinking mode is used.
Is Qwen4-Max available as open weights?
The Max variant is hosted only through Alibaba APIs. Smaller open-weight Qwen4 models are planned but have no confirmed release date.
What are the current API prices for Qwen4-Max?
Pricing is set at 1.50 dollars per million input tokens and 4.50 dollars per million output tokens, with cached input at 0.075 dollars per million.
Comments
No comments yet. Be the first!