Qwen4-Max: Alibaba Bets on Workforce Agents

With Qwen4-Max, Alibaba targets industrial, scientific, and financial applications in competition with Opus 5.5, GPT-6, and Kimi K3.1 Pro. Artificial Analysis indicates a lower cost per task compared to its predecessor, but results against competitors vary depending on the tests.
Alibaba Claims 48-Hour Autonomous Experiments
Alibaba claims that a model from the Qwen generation has conducted physical manipulations for 48 hours without intervention, in conjunction with Mira, the system from the startup Deep Principle. The experiment focused on the design, synthesis, and evaluation of a liquid intended for cooling high-performance data center chips. According to the company, 2.7 million molecular candidates were explored, while an automated laboratory executed four cycles combining design, synthesis, and measurement.
In pharmaceutical research, Alibaba reports a collaboration between Qwen4-Max and PharmMaster, a platform from DP Technology, a Beijing-based startup specializing in AI drug discovery. The target was an inflammatory response associated with Alzheimer’s: literature searches, calculations, and design work led to the optimization of 24 candidate molecules. Alibaba attributes to the selected candidate a binding to the target protein approximately 3.6 times better than that of the initial compound described in a patent. DP Technology specialists consider these results encouraging enough to initiate new laboratory trials.
Live Delivery Allows Intervention Before Final Result
Alibaba, like much of the industry, organizes this version around the autonomous execution of long tasks. The company responds with Live Delivery to a challenge: waiting for the final delivery leaves little room for correcting errors. A usable version appears quickly, then is enriched over iterations; the user can review intermediate outputs and redirect the task during its execution. The Qwen-Live-Delivery-Plugin module has been made available as open source.
Alibaba also showcases its model with Unity, Unreal Engine, or Godot to produce a playable game from a single description. The claimed scope ranges from scene organization to game logic, including external assets, materials, interfaces, and visual effects. The example Meow of Magic, created in Unreal Engine, features a wizard cat casting spells against waves of opponents and bosses.
According to Alibaba, the model executes the game itself during its creation and uses interactive and visual feedback to refine the camera, animations, impact reactions, and lighting. The company also attributes the creation of reusable 3D objects in Blender for use in games or for 3D printing.
Alibaba Quantifies Industrial Gains on Work360
To evaluate industrial work, Alibaba uses its own benchmark, Work360, covering 721 specialties. The company announces a token consumption lower by about a quarter to a third to accomplish the same tasks as the previous model. The average reported cost decreases by 62%, from $4.43 to $1.67, while the score improves from 47.74 to 52.69. On Terminal-Bench 4.0, Alibaba reports a token consumption nearly halved.
The deployments mentioned are also results communicated by Alibaba. At automotive partners SANF and Pacific Union, the manual programming of a part reportedly took 24 minutes compared to 475 minutes previously. For finance, the company describes the use of 90 MCP tools collecting market data from various sources. The deliverables are verifiable Excel and PowerPoint files, with each figure linked to its source.
Alibaba also employs Qwen4-Max in its own cloud infrastructure. Paired with the internal solver MindOpt, it reportedly reduced the average cost of allocating calculations dedicated to data annotation by 10.42%. In the distribution of CPU tasks in data centers, the company announces optimal solutions obtained more than 14 times faster than with the comparison method.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Artificial Analysis Estimates Task Cost at $3.27
Artificial Analysis estimates the cost of a task from its Intelligence Index for Qwen4-Max at $3.27, compared to $10.02 for Qwen3.8-Max, which is about three times higher. The estimate for Kimi K3.1 Pro reaches $3.68, positioning Qwen4-Max at a level approximately 11% lower. This cost reduction combines a 25% decrease in token pricing, 70% cheaper cache access, and reduced usage: the average number of output tokens per task in the benchmark is around 105,000, down from 171,000 previously. Tencent Hy4 shows an even lower rate, at $1.03 per task, but achieves an eight-point lower score on the index.
At the maximum reasoning level, Qwen4-Max reaches 43 points, nine above Qwen3.8-Max and three below Kimi K3.1 Pro, the first Chinese model with 46 points. Progressions on HLE-Diamond reach 21 points, resulting in 45%, compared to gains of 18 points on AA-Omniscience and 17 on AutomationBench-AA. Launched in October by Moonshot AI, Kimi K3.1 Pro has 2.8 trillion parameters. On Terminal-Bench 4.0, it scores 67% compared to 48% for Qwen4-Max; on Cognition's FrontierCode 2.0, both reach 26%, behind Tencent Hy4 at 31%.
Qwen4-Max tops the compared group on Agents' Last Exam v1.1 with 46.7 points. On Humanity's Last Exam, it falls behind Opus 5.5 and the GPT models; on HealthBench Professional, it surpasses Opus 5.5 but not the GPT models. Alibaba claims an advantage for agents and professional tasks while acknowledging the results of Opus 5.5 and the GPT models on Humanity's Last Exam.
Qwen4-Max Activates 71 Billion Parameters Per Token
According to the Qwen team, the mixture-of-experts architecture of Qwen4-Max comprises 2.4 trillion parameters, with 71 billion activated for each token. The model natively supports images and accepts a context of one million tokens. Its predecessor Qwen3.8-Max, launched in August for autonomous tasks over several days, also had 2.4 trillion parameters but activated 95 billion per token. Artificial Analysis had measured it at 64 steps per task compared to 14 for Qwen3.7-Max, with a cost more than twice as high.
Qwen4-Max introduces a revision of the Qwen4 architecture: three layers condense previous information, followed by a fourth that retains important elements. According to Alibaba, this differentiated attention avoids uniform processing of context and reduces computational needs on long inputs. Four parallel connections between layers, equipped with dynamic gates, aim to stabilize the flow of information.
An N-gram embedding layer containing 328 billion parameters provides frequent sequences of already formed tokens. Alibaba claims that training has been stable, with loss decreasing steadily, without sharp spikes. Introduced at the end of August with Qwen3.8-Flash-Next, the Qwen4 architecture already included an N-gram embedding layer with 51 billion parameters.
Alibaba Reserves Open Weights for Smaller Qwen4 Models
Alibaba launched Qwen4-Max for industrial, scientific, and financial professional uses, particularly against Opus 5.5, GPT-6, and Kimi K3.1 Pro. The model is accessible on QwenCloud and in the QwenWork, Qoder, and QwenStudio products. On its API, Alibaba charges $1.50 per million tokens for input and $4.50 for output; the indicated rate for cache access is $0.075.
The company also claims safer code: on Last Secure Code Bench, it announces that the proportion of outputs that are both functional and secure reaches 31.8%, up from 22% previously. The share deemed safe at each test level rises from 8% to 14.7%, which Alibaba presents as nearly a doubling. For CBRN risks, covering chemical, biological, radiological, and nuclear domains, the company indicates that its internal reviewers have scrutinized all thought chains of the model. They reportedly found no effective innovation capability that would give a significant advantage to a user in these areas.
The weights of Qwen3.8-Max were published in mid-August. For the Qwen4 generation, Alibaba plans to later release smaller versions as open source; the opening of weights currently only concerns these models.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.