Back Home

最新模型

Alibaba Launches Qwen4-Max: 71 Billion Parameters Activated Per Token, with a Million-Token Context

Alibaba released Qwen4-Max on November 16, featuring an updated MoE architecture and native vision capabilities. The production API also offers explicit caching and reasoning summaries, making cost and state management key considerations for long-running agent tasks.

Thomas LOMBARD, designed by HASSELL (architects)[1] · CC BY-SA 3.0 · Image source
zh-Hant

Alibaba released Qwen4-Max on November 16, offering Token Plan and API access through QwenCloud. The company lists 2.4 trillion total parameters, with 71 billion activated per token—25% fewer than Qwen3.8-Max, which has the same 2.4 trillion-parameter scale. The updated sparse MoE architecture revises attention, residual connections, and embedding design, and supports native vision and a million-token context. Official announcement

For developers, the specific changes include the production model ID qwen4-max, along with support for text, image, and video inputs and tool calling. The documentation lists a maximum output of 131,072 tokens in non-reasoning mode, and a combined limit of 262,144 tokens for reasoning and the final answer. High-intensity reasoning is enabled by default, so applications should set output budgets explicitly to avoid unexpected usage on long tasks. Model documentation

Agent integrations also gain explicitly configurable caching: shared system prompts, tool definitions, and reference documents can be assigned cache breakpoints, with a time-to-live of five minutes or one hour. CoT Summary separates a readable reasoning summary from encrypted full reasoning; the documentation requires subsequent turns to return the encrypted content unchanged, and the summary itself cannot replace the reasoning history. This makes message retention for multi-turn tool calls an integration detail to check during migration. Integration guide

In Singapore, prices per million input and output tokens are $1.50 and $4.50, respectively. Reasoning tokens are billed at the output rate, while cache reads cost $0.075. Explicit cache writes are billed separately, so engineering teams should estimate total costs based on prefix reuse frequency and cache duration, and include tool fees in task billing. Pricing documentation

On performance, Alibaba claims that in its own Work360 benchmark, high-intensity reasoning used 35% fewer tokens than the previous generation and reduced average cost per task by about 60%. These are vendor test results and cannot be directly extrapolated to other agent workloads. Independent task success rates, long-context latency, and real-world costs remain worth watching. The current documentation explicitly lists batch inference, fine-tuning, and model deployment as unsupported. Alibaba has also previewed a smaller open-weight Qwen4 release, but has not announced its size or release date. Launch announcement, feature limitations

Sources

  1. qwen4-max-preview — Alibaba Cloud Model Studio
  2. Latest model: Qwen4-Max-Preview
  3. API compatibility upgrades
  4. Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery
  5. Pricing — QwenCloud