Alibaba’s Qwen4-Max Pushes Long-Context AI Into New Territory!

Alibaba has opened a new chapter for its Qwen model family with Qwen4-Max, a giant multimodal system built to handle long-horizon professional work. The first model in the Qwen4 generation combines a 2.4T-parameter backbone with 71B parameters active per token, accepts text, images, and video, and supports a 1M-token context window.
That scale arrives with a second headline: Alibaba reports a 62% drop in average cost per task, alongside lower API prices. Qwen4-Max also brings strict cybersecurity guardrails, showing how the race to build more capable AI now runs alongside a race to control what those systems can do.
A 2.4T-Parameter Model With Focused Computation
Qwen4-Max is a proprietary sparse mixture-of-experts model. Its 2.4T-parameter backbone gives the system a huge pool of learned capacity, but only 71B parameters activate for each token. That design aims to deliver the reach of a massive model without using the full parameter count for every step.
The Qwen4 team upgraded its sparse MoE design in three places: attention, residual connections, and embeddings. Each QSA layer pairs with three Gated DeltaNet layers, which compress history into fixed-size states. QSA then uses a lightweight indexer to select relevant context at micro-block granularity, lowering attention and indexing costs across long sequences.
That architecture matters when a model must work through huge collections of information instead of a short prompt. The 1M-token context window gives Qwen4-Max room to process long professional tasks, while support for text, images, and video expands the material it can examine.
The model also adds Gated Residual, or GR, with four parallel residual branches and dynamic read and write gates. N-gram Embedding adds capacity with little extra compute, using 328B extra parameters to look up local token patterns at layer 2.
Alibaba reports no loss spikes across the 2.4T-scale pre-training run. On in-house tests, Qwen4-Max-Base delivered 5 to 13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters.
More Work Per Token, Less Cost Per Task
The performance story centers on professional workloads. Alibaba’s comparison chart covers 16 benchmarks, and Qwen4-Max ranks first on 4 of them. The model trails most clearly on coding and vision, but its results on Alibaba’s Work360 benchmark show a strong gain across broad professional tasks.
Work360 spans 26 top-level domains and 721 specialized fields. Qwen4-Max scores 52.69, up from 47.74 for Qwen3.8-Max. Token use drops about 24% at Xhigh effort and 35% at High effort on Work360, while Terminal-Bench 4.0 shows drops of about 48% and 49%.
Those reductions translate into a lower bill. Average Work360 cost per task falls from $4.43 to $1.67, which Alibaba describes as a 62% cut. The change comes as the model also reduces its published API prices:
- Input costs $1.50 per million tokens, down from $2 for Qwen3.8-Max.
- Output costs $4.50, down from $6.
- Implicit cache reads fall 70%, from $0.25 to $0.075.
The model supports up to 393K max output tokens, giving long-running tasks room to produce substantial results. On QwenCloud, its model ID is qwen4-max, and the API works with OpenAI-compatible and Anthropic-compatible clients.
Developers can use function calling, structured outputs, context caching, batch processing, web search, and fine-tuning. The default rate limit is 5,000,000 tokens per minute, while thinking mode can be toggled with enable_thinking.
Access runs through the Token Plan, and a Token Plan-only preview opened this month only. Qwen Studio, QwenWork, and Qoder are listed as run environments for Qwen4-Max, giving users several paths into the same model family.
Capability Meets a New Cybersecurity Boundary
Alibaba Group Holding calls Qwen4-Max its most advanced artificial intelligence model to date, but the company is pairing that capability with limits on offensive use. Alibaba said it intentionally filtered the model’s training data to remove offensive cyberattack capabilities while retaining cyber defence skills such as vulnerability detection.
The company also introduced dedicated data sets and post-training enhancements to mitigate chemical, biological, radiological, and nuclear risks. That safety work places Qwen4-Max inside a wider shift among China’s leading artificial intelligence model companies.
Tencent Holdings said it stripped its latest Hy4 model of cyber-hacking capabilities while retaining cyber defence capabilities. Moonshot AI made a similar move with its Kimi K3.1 Pro model in October.
The challenge becomes sharper when models are released as open-weight software. Anyone with sufficient hardware can download and modify those models, including malicious actors. Qwen4-Max is described as proprietary, with hosted access through QwenCloud and subscription access through the Token Plan, but its guardrails reflect the same pressure facing the wider AI industry: more capability creates more value, and more risk.
Qwen4-Max arrives with a clear ambition. It aims to stretch context windows, process multiple forms of media, reduce token use, and make demanding professional work cheaper to run. Its next test will be whether those architectural gains and cybersecurity boundaries hold as developers put the 1M-token system to work.
Based on
- Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context — marktechpost.com
- Alibaba says new flagship AI model cuts task costs by 60% while tightening cyber guardrails | South China Morning Post — scmp.com
- Can Chinese AI companies curb cyber risks from open-weight models without losing their edge? | South China Morning Post — scmp.com
- Adventures coding with AI | Page 81 | Ars OpenForum — arstechnica.com




