Alibaba launches Qwen4-Max, a 2.4T-parameter model at lower cost per task

Alibaba launches Qwen4-Max, a 2.4T-parameter model at lower cost per task

Alibaba has launched Qwen4-Max, which it pitches as its answer to Opus 5.5, GPT-6 and Kimi K3.1 Pro. The mixture-of-experts model has 2.4 trillion parameters but activates only 71 billion per token, according to the Qwen team. It processes images natively and offers a one-million-token context window. It targets professional work in industry, science and finance.

The model debuts the revamped architecture of the Qwen4 series. Instead of relying only on standard attention layers, Alibaba combines two methods: three layers compress prior context into a condensed state, and a fourth layer picks out only the parts that matter. That sharply cuts compute on long inputs. The design also adds four parallel connections between layers, with dynamic gates meant to stabilize information flow, and an N-gram embedding layer with 328 billion parameters that supplies common token sequences as ready-made building blocks. Alibaba says training the 2.4-trillion-parameter configuration ran stably, with training loss falling steadily and no sudden spikes. In Alibaba's internal bits-per-byte evaluations, Qwen4-Max-Base scored best among five checkpoints and improved 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.

Efficiency shows up in running costs. On Work360, Alibaba's own benchmark for industrial tasks across 721 specialized fields, Qwen4-Max uses roughly a quarter to a third fewer tokens than its predecessor, and on the agent benchmark Terminal-Bench 4.0 token use drops by nearly half. Alibaba says the average cost per Work360 task falls 62 percent, from $4.43 to $1.67, while the score rises from 47.74 to 52.69.

External results are mixed. On Agents' Last Exam v1.1, Qwen4-Max scores 46.7 points, the best in the comparison group. It trails Opus 5.5 and the GPT models on Humanity's Last Exam. On HealthBench Professional it falls behind the GPT models but edges past Opus 5.5.

On the Artificial Analysis Intelligence Index, at its highest reasoning setting, the model scores 43 points: nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro, the top Chinese model at 46. The biggest gains over the predecessor are on HLE-Diamond (up 21 points to 45 percent), AA-Omniscience (up 18) and AutomationBench-AA (up 17). Artificial Analysis puts the cost at $3.27 per Intelligence Index task, about a third of Qwen3.8-Max's $10.02 and roughly 11 percent less than Kimi K3.1 Pro's $3.68. The savings come from three places: token prices 25 percent lower than the predecessor's, cache hits costing 70 percent less, and fewer tokens used (about 105,000 output tokens per benchmark task against 171,000 for Qwen3.8-Max). Tencent's Hy4 is much cheaper at $1.03 per task but scores eight points lower on the index. On Terminal-Bench 4.0, Qwen4-Max scores 48 percent, behind Kimi K3.1 Pro at 67 percent. On Cognition's FrontierCode 2.0 it ties Kimi at 26 percent and trails Tencent Hy4 at 31 percent. Alibaba's own API charges $1.50 per million input tokens and $4.50 per million output tokens, with cache hits at $0.075.

The release is built around agents that handle long tasks on their own. Alibaba's feature for this is Live Delivery: the model delivers a usable first version early and builds it out in later passes, and users can step in and change course at any point. Alibaba has open-sourced the accompanying Qwen-Live-Delivery-Plugin.

Alibaba cites several deployments. Automotive partners SANF and Pacific Union reportedly used Qwen4-Max to cut manual programming time for a single workpiece from about 475 minutes to 24. In finance, the model taps 90 MCP tools to pull market data from multiple sources and delivers verifiable Excel and PowerPoint files in which every number traces back to its source. Internally, Alibaba says Qwen4-Max worked with its in-house solver MindOpt to cut the cost of allocating compute jobs for data annotation by an average of 10.42 percent, and found optimal solutions more than 14 times faster than a comparison method when distributing CPU tasks in data centers.

Alibaba also shows the model in physical labs. Paired with Mira, a system from startup Deep Principle, it ran autonomously for 48 hours, according to Alibaba, and designed, synthesized and tested a new coolant for high-performance data center chips. It searched 2.7 million molecular candidates, controlled an automated lab and completed four full cycles of design, synthesis and measurement. With PharmMaster, a platform from Beijing-based AI drug discovery startup DP Technology, it worked on a compound aimed at an inflammatory response linked to Alzheimer's disease. Over several rounds it optimized 24 candidate molecules; Alibaba says the final candidate binds to the target protein about 3.6 times better than the known starting compound from a patent. DP Technology's experts consider the result promising enough to pursue further lab tests.

Alibaba also shows the model working with Unity, Unreal Engine and Godot. From a single natural-language description it is supposed to handle the whole development process, from scene setup and game logic to assets, materials, UI and visual effects, until the game is playable. It runs the game itself and uses visual and interactive feedback to adjust camera work, animations, hit reactions and lighting. The sample project is "Meow of Magic", an Unreal Engine adventure about a cat wizard. Alibaba says the model also builds 3D models in Blender for games and 3D printing. The article notes that Anthropic moved Claude Code the opposite way, barring it from recreating existing games after a wave of unofficial Halo and GTA: Vice City clones.

On safety, Alibaba says the share of code that is both functional and secure on Last Secure Code Bench rose from 22 to 31.8 percent, and the share rated secure at all three testing levels nearly doubled, from 8 to 14.7 percent. For CBRN risks (chemical, biological, radiological, nuclear), internal reviewers went through the model's full chains of thought, the company says, and found no real capacity for innovation that would give a user a significant boost in capability.

Qwen4-Max is available through QwenCloud and in QwenWork, Qoder and QwenStudio. Alibaba plans to release smaller Qwen4 models as open source later; for Qwen4, only the smaller models will get open weights for now. For context, Alibaba released Qwen3.8-Max in August, its first Max model built for autonomous tasks that run for days, also with 2.4 trillion parameters but 95 billion active per token. Moonshot AI followed in October with Kimi K3.1 Pro at 2.8 trillion parameters.

Key facts

  • Qwen4-Max is a mixture-of-experts model with 2.4 trillion parameters, 71 billion active per token, native image input and a one-million-token context window.
  • On the Artificial Analysis Intelligence Index it scores 43 at its highest reasoning setting: nine more than Qwen3.8-Max, three behind Kimi K3.1 Pro (46).
  • Artificial Analysis puts its cost at $3.27 per index task, about a third of Qwen3.8-Max's $10.02 and roughly 11 percent below Kimi K3.1 Pro's $3.68.
  • Alibaba's API costs $1.50 per million input tokens and $4.50 per million output tokens; cache hits cost $0.075.
  • Live Delivery lets users inspect early results from long-running agents and change course; Alibaba open-sourced a plugin for it.

Why it matters

Qwen4-Max is Alibaba's new flagship and the first model to use the revamped Qwen4 architecture, which mixes compressing layers with a selective layer to make long contexts cheaper. The headline result is cost: Artificial Analysis measures $3.27 per Intelligence Index task, about a third of the predecessor's $10.02 and roughly 11 percent below Kimi K3.1 Pro. The model does not lead the field. It is three index points behind Kimi K3.1 Pro, trails Opus 5.5 and the GPT models on Humanity's Last Exam, and is well behind Kimi on Terminal-Bench 4.0 (48 versus 67 percent).

Who it affects

Teams running long, agentic workloads in industry, science and finance are the stated target, along with developers who pay per token. Alibaba names partners and collaborators in its examples: SANF and Pacific Union in automotive, Deep Principle and DP Technology in lab and drug-discovery work. Rival labs are in the frame too: Moonshot AI's Kimi K3.1 Pro, Tencent's Hy4 (cheaper at $1.03 per task but eight index points lower), Opus 5.5 and GPT-6.

How to use it

Qwen4-Max is available through QwenCloud and in QwenWork, Qoder and QwenStudio. Alibaba's own API charges $1.50 per million input tokens and $4.50 per million output tokens, with cache hits at $0.075. Token prices are 25 percent lower than the predecessor's. Alibaba has open-sourced the Qwen-Live-Delivery-Plugin for the Live Delivery feature. For Qwen4, only the smaller models will get open weights for now, and Alibaba plans to release them later.

How solid is it

The index score and the cost per task come from Artificial Analysis, an independent benchmarker, so they are the firmest numbers here. Most other claims are Alibaba's own: the Work360 results (a benchmark Alibaba built, with the 62 percent cost cut), the internal bits-per-byte evaluations, the partner deployments, the lab and drug-discovery experiments, the game-development demos and the safety figures. The article reports no independent verification of those. The Work360 cost cut is a different figure from the roughly one-third cost on the Artificial Analysis index. External benchmark results are described as mixed.

Risks and caveats

The game-building ability is described as something the model is "supposed to" do from a single description, and the demo is a vendor sample project. The CBRN assessment rests on Alibaba's internal reviewers reading the model's chains of thought. The drug-discovery result is a 3.6-fold binding improvement over a known starting compound, per Alibaba, and DP Technology's experts say it justifies further lab tests, not that it is proven. The cost lead is also narrow against Kimi K3.1 Pro (about 11 percent) and does not extend to Tencent's Hy4, which is cheaper still.