Grafy News
Try Grafy free

Alibaba launches Qwen4-Max: 2.4 T‑parameter MoE model with 1 M‑token context

Alibaba unveiled Qwen4-Max, a 2.4 trillion‑parameter Mixture‑of‑Experts model with 71 billion active weights.

Illustration for the news story: Alibaba launches Qwen4-Max: 2.4 T‑parameter MoE model with 1 M‑token contextSource: MarkTechPost

Alibaba announced the release of Qwen4‑Max, a proprietary large‑scale model that combines a 2.4 trillion‑parameter Mixture‑of‑Experts (MoE) backbone with 71 billion actively used parameters, native vision capabilities, and a context window of up to 1 million tokens. The announcement matters because it pushes the frontier of both model size and practical context length, while also offering a pricing structure that could reshape cost calculations for developers and researchers.

What is Qwen4‑Max and how is it built?

According to MarkTechPost, Qwen4‑Max follows the MoE paradigm, where a massive pool of parameters is sparsely activated per token. The total parameter count reaches 2.42.4 trillion, but only 7171 billion are engaged during inference, keeping latency manageable. The model also integrates vision encoders natively, allowing it to process images alongside text without a separate multimodal wrapper. This design mirrors earlier Alibaba releases such as Qwen‑2, but scales the expert routing and vision modules substantially.

How does the 1‑million‑token context change what can be done?

A context window of 1,000,0001{,}000{,}000 tokens is an order of magnitude larger than the typical 8,0008{,}000‑to‑32,00032{,}000 token windows found in most commercial LLMs. In practice, this enables:

  • Long‑form document analysis where the entire source can be fed at once, removing the need for chunking and re‑assembly.
  • Complex codebases or data tables that exceed conventional limits, allowing a single prompt to reference all relevant sections.
  • Extended conversational memory for chat agents, preserving nuanced history over many interactions. The sheer length also opens new research avenues in hierarchical reasoning and retrieval‑augmented generation, where the model can directly attend to far‑flung evidence.

What are the cost and efficiency claims?

MarkTechPost reports that Qwen4‑Max is priced at 1.51.5 per million input tokens, a figure that is positioned as competitive for a model of this scale. Additionally, Alibaba claims a 62 % reduction in the per‑task cost for its Work360 platform when using Qwen4‑Max compared with prior offerings. The lower token price, combined with the ability to handle longer contexts, can translate into fewer API calls and reduced overall expense for workloads that previously required multiple passes.

How does Qwen4‑Max compare to existing models?

While the article does not provide direct benchmark numbers, the specifications suggest that Qwen4‑Max sits alongside other frontier models such as GPT‑4‑Turbo and Claude‑3, which typically expose around 100100 billion active parameters and context windows under 100,000100{,}000 tokens. The MoE architecture gives Qwen4‑Max a higher total parameter budget without proportionally higher inference cost, a trade‑off that has been shown to improve few‑shot performance on many benchmarks. Its native vision support also differentiates it from text‑only giants, providing a single endpoint for multimodal tasks.

How can creators and developers try it today?

Alibaba makes Qwen4‑Max available through its cloud AI marketplace, where developers can obtain API keys and select the pricing tier that matches their token usage. Integration follows the usual REST or gRPC patterns, with request payloads accepting both text and image data. For teams already using Grafy, the model can be routed through the Grafy Chat interface, which lets users experiment with multiple models side‑by‑side. By adding Qwen4‑Max as a custom endpoint in Grafy Chat, creators can compare its responses against other LLMs and immediately see cost implications.

Grafy perspective: exploring Qwen4‑Max with Grafy

  • Chat – Use the Grafy Chat app to add Qwen4‑Max as a selectable model, enabling rapid side‑by‑side comparisons with existing models in the same workspace.
  • Code – The Grafy Code studio can generate request wrappers in Python, JavaScript or other languages, letting developers prototype API calls, batch token calculations, and error handling without leaving the platform.
  • Reader – For researchers examining the model’s technical paper or related documentation, the Grafy Reader app offers AI‑assisted search and question‑answering over PDFs, making it easier to extract implementation details or benchmark methodologies. If a direct integration is not yet available, the Grafy News Desk recommends using Grafy Chat’s “lowest‑cost model routing” pattern (see the blog post on that topic) to automatically fall back to cheaper models when token budgets are tight, while still testing Qwen4‑Max on high‑value prompts.

Frequently asked questions

What does the "71 B active parameters" figure mean?

In a Mixture‑of‑Experts model, the total pool of parameters is larger than the subset used for any single token. Qwen4‑Max has 2.42.4 trillion parameters overall, but only 7171 billion are activated during inference, keeping compute and latency comparable to dense models of that size.

How is the 1.51.5 per‑million‑token price calculated?

The price refers to input tokens only; output tokens are billed separately according to Alibaba’s standard rates. For a prompt that consumes 10,00010{,}000 tokens, the cost would be 0.0150.015, not including any additional output token charges.

Can I use Qwen4‑Max for image generation?

Qwen4‑Max includes a native vision encoder, allowing it to accept image inputs alongside text. However, it is not an image generation model; it can analyze or describe images but does not produce new visual content.


Source: MarkTechPost — Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context