Alibaba’s Qwen team released Qwen4-Max on 16 November and says it lowers the average cost of a task from $4.43 to $1.67 compared with its previous flagship, Qwen3.8-Max. That is a 62 percent drop. The number comes from Work360, a benchmark Qwen designed and runs itself, and the announcement includes no independent results.

Under the hood, Qwen4-Max routes each token through experts: the backbone totals 2.4 trillion parameters, 71 billion of them active per token, with 328 billion N-gram Embedding parameters on top. Vision is native, and the context window reaches one million tokens. Qwen says the training run at this scale had no loss spikes. In its in-house tests, the Qwen4-Max-Base checkpoint posted bits per byte 5 to 13 percent lower than Qwen3.8-Max-Base while activating 25 percent fewer parameters.

The savings come from two places. Qwen says token usage on Work360 falls about 24 percent at Xhigh effort and 35 percent at High versus Qwen3.8-Max, and about 48 and 49 percent on Terminal-Bench 4.0. API prices are also lower than the predecessor’s, with large cuts to cached-input pricing. On quality, Qwen4-Max scored 52.69 on Work360 against 47.74 for the older model. Work360 sorts work into 26 top-level domains, 182 subdomains and 721 fields, and Qwen wrote the taxonomy and the scoring.

The showcase list is wide. Qwen says the model builds 3D scenes in Blender, covering geometry, materials, lighting and rendering. It also says the model produced playable games in Unity, Unreal Engine and Godot from one written brief, including an installable Android racer set around West Lake. At the automakers SANF and Pacific Union, Qwen reports that programming one workpiece by hand used to take about 475 minutes and now takes 24.

Professional work gets the same treatment. In a finance task, Qwen says the model called 90 MCP tools and handed back Excel workbooks plus PowerPoint decks that can be edited, each tracing to its sources. In pre-IPO legal due diligence, it sorted 198 inconsistently named working papers into 14 folders and wrote an 86-page report flagging 14 legal issues. In critical care, it reviewed a patient record that included 1,433 original medical images and proposed four care recommendations. Qwen says they agreed with what the physician really ordered. Each example is a single case that Qwen chose to show.

The science claims run longest. On Deep Principle’s Mira platform, an agent ran for 48 hours straight, generating 2.7 million candidate molecules. After four rounds with an automated wet lab, Qwen says it landed on one coolant molecule, confirmed by experiment, meant for cooling chips in data centers. On DP Technology’s PharmMaster, aimed at the NLRP3 target tied to Alzheimer’s neuroinflammation, the model produced 24 candidate molecules, and Qwen puts the best computed binding affinity near 3.6 times that of the baseline compound. Qwen notes those still need wet-lab validation.

Security results are more modest. On Last Secure Code Bench, the function-level pass rate for code that is both functional and secure rose from 22.0 percent to 31.8 percent. The stricter test, secure code across function, file and repository at once, moved from 8 percent to 14.7 percent. Even the better score means roughly two in three generated functions fail on function or security. Qwen says its internal review saw nothing that would enable large-scale cyberattacks and found no CBRN uplift risk.

Qwen4-Max is available through QwenCloud (by Token Plan or API) and inside QwenWork, Qoder and QwenStudio. Qwen is open-sourcing its Qwen-Live-Delivery-Plugin, which supports long-running agents that ship an early version and then release updates while users step in. The company also says smaller models in the family are coming to open source, with no dates given.

Cheaper tokens matter most to teams running agents all day, where cost scales with every tool call. Because Work360 is Qwen’s own mix of tasks, the $1.67 figure is a claim to test: run your own workload on both Qwen4-Max and Qwen3.8-Max before budgeting around it.

Qwen Team, Alibaba, qwen.ai blog, 16 November 2026.