Alibaba's Qwen team introduces Qwen4-Max with a 2.4 trillion parameter backbone

Alibaba's Qwen team introduced Qwen4-Max on 16 November 2026, describing it as a new AI teammate for work, creativity and discovery. Its context window runs to one million tokens, and the team says the model has native vision and a backbone of 2.4 trillion parameters. QwenCloud, where the model is offered, lists input at $1.5 per million tokens and output at $4.5 per million tokens.
Architecture and training results
According to the Qwen team, the Qwen4 series uses a new architecture in which 71 billion parameters are activated per token. The design also carries an extra 328 billion N-gram Embedding parameters beyond the backbone. The N-gram component performs context-based lookups at the second layer, which the team says expands capacity with little additional computation. The team says the sparse MoE model is changed in three places: attention, residual connections and embeddings. Training at that scale went smoothly, and the team reports no loss spikes throughout the run. The team attributes this stability mainly to improvements in model architecture and training optimization. In the team's in-house evaluations, Qwen4-Max-Base reached the lowest bits per byte among the five checkpoints evaluated. Against Qwen3.8-Max-Base, the new base model improved in every domain, with roughly 5 to 13 percent lower BPB while activating 25 percent fewer parameters.
Token use, cost and benchmark scores
On the team's Work360 evaluation, the average cost per task fell from $4.43 to $1.67, which the blog calls a 62 percent reduction. Token use on Terminal-Bench 4.0 fell by approximately 48 percent at Xhigh effort and 49 percent at High effort, according to the same post. The team also says it lowered API prices against Qwen3.8-Max, with the largest cuts in cached-input pricing. On Work360, a benchmark the team built for industry tasks, Qwen4-Max scored 52.69 against 47.74 for Qwen3.8-Max. Work360 sorts work into 26 top-level domains, 182 subdomains and 721 specialized fields, the blog says. The QwenCloud model page lists a maximum input of 991K tokens and a maximum output of 393K tokens.
Industry and research examples
The blog names healthcare, law and finance as target fields, along with game development, 3D modeling, precision manufacturing and frontier scientific discovery. In a showcase at SANF and Pacific Union, Qwen4-Max cut the manual effort to program one workpiece from roughly 475 minutes to 24 minutes, a drop the blog puts at about 95 percent. In a pre-IPO legal due diligence example, the model sorted 198 inconsistently formatted working papers into 14 folders and compiled an index, according to the blog. In the finance example, the team says Qwen4 invoked 90 MCP tools to gather market quotes, financials, interest rates and dividend data.
Long-running agents and laboratory work
The post describes a Live Delivery feature in which the agent hands over a usable first version early and then improves it through successive releases. The team is also open-sourcing the Qwen-Live-Delivery-Plugin for users who want to try the feature. In a continuous 48-hour agent run on Deep Principle's Mira platform, the blog says Qwen4 built a candidate space of 2.7 million molecules. After four iterations of a closed design, synthesis and learning loop, Qwen4 obtained an experimentally validated coolant molecule, the blog says. On DP Technology's PharmMaster platform, Qwen4-Max made 454 tool calls, ran 3,239 computational metric evaluations and designed and optimized 24 candidate molecules. The blog reports endorsement from DP Technology, which sees substantial research value in the resulting molecule and says it warrants further wet-lab validation.
Availability, plans and what comes next
Qwen4-Max is available on QwenCloud, and users can start with either the Token Plan or the API, the blog says. The post also names QwenWork, Qoder and QwenStudio as other places to use the model. QwenCloud says first-time customers can pay as little as $3 for a Token Plan during an event, with the final amount fixed at checkout after login. The Token Plan page says the Essential plan carries 2.25 times the credits of the Lite plan. The Qwen team says smaller models in the Qwen4 family are coming to open source, tuned for what small models do best.
