AIsaAIsaStudio 中文 Open AIsa Studio

Alibaba's Qwen4-Max takes on Opus 5.5, GPT-6, and Kimi K3.1 Pro at much lower prices

Alibaba released its new flagship Qwen4-Max, a 2.4-trillion-parameter MoE model with 71 billion active parameters per token and a one-million-token context window. Token use on agent benchmarks fell by roughly a quarter to half versus its predecessor, with per-task cost on Alibaba's Work360 benchmark dropping from $4.43 to $1.67, and API pricing set at $1.50 per million input and $4.50 per million output tokens. The company also demonstrated the model running a lab autonomously for 48 hours, optimizing drug candidate molecules, and building playable games in engines like Unreal Engine, while external benchmark results are mixed against Opus 5.5 and the GPT models.

Why it matters

The piece compiles architecture details, external benchmark results, and pricing data, including where the model trails rivals, letting readers weigh its cost-performance trade-offs.

Read the original ↗AI Daily, Nov 17

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.

Analysis

What happened

Alibaba released Qwen4-Max, a new flagship model positioned as its answer to Opus 5.5, GPT-6, and Kimi K3.1 Pro. The company pitches it as pairing a more efficient architecture with much lower prices, targeting professional work in industry, science, and finance.

The model is a mixture-of-experts system with 2.4 trillion parameters, of which 71 billion are activated per token. It processes images natively and offers a one-million-token context window. It is available through QwenCloud and in QwenWork, Qoder, and QwenStudio, and Alibaba plans to release smaller Qwen4 models as open source later.

Key details

  • Hybrid attention architecture: Instead of relying only on standard attention layers, three layers compress prior context into a condensed state and a fourth layer picks out only the relevant parts, sharply cutting compute costs on long inputs. The architecture also adds four parallel connections between layers with dynamic gates, plus an N-gram embedding layer with 328 billion parameters that supplies common token sequences as ready-made building blocks.
  • Training stability: According to Alibaba, training the 2.4-trillion-parameter configuration ran stably, with training loss declining steadily and no sudden spikes. In internal evaluations measured in bits per byte across knowledge, code, question answering, and STEM, 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.
  • Lower token use: On Alibaba's Work360 benchmark (721 specialized fields), the model uses roughly a quarter to a third fewer tokens than its predecessor; on Terminal-Bench 4.0, token use drops by nearly half. Average cost per Work360 task falls 62 percent, from $4.43 to $1.67, while the benchmark score climbs from 47.74 to 52.69.
  • Mixed external results: Qwen4-Max scores 46.7 on Agents' Last Exam v1.1, the best in its comparison group, but trails Opus 5.5 and the GPT models on Humanity's Last Exam and the GPT models on HealthBench Professional. It ties Kimi K3.1 Pro at 26 percent on FrontierCode 2.0 and trails Tencent Hy4 at 31 percent.
  • Pricing: Via its API, Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens, with cache hits at $0.075. 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. Savings come from 25 percent lower token prices, 70 percent cheaper cache hits, and fewer tokens consumed (about 105,000 output tokens per benchmark task versus 171,000 for the predecessor).
  • Agent features and deployments: The release introduces "Live Delivery," where the model delivers a usable first version early and builds it out in later passes, letting users change course at any point; the accompanying Qwen-Live-Delivery-Plugin is open-sourced. Alibaba cites automotive partners SANF and Pacific Union cutting manual programming time for a single workpiece from about 475 minutes to 24, and in finance the model taps 90 MCP tools to produce verifiable Excel and PowerPoint files. Internally, it worked with the solver MindOpt to cut data annotation compute allocation costs by an average of 10.42 percent.
  • Science and gaming demos: Paired with Deep Principle's Mira system, the model ran autonomously for 48 hours, searching 2.7 million molecular candidates to design, synthesize, and test a new coolant for data center chips across four full cycles. With DP Technology's PharmMaster, it optimized 24 candidate molecules targeting an inflammatory response linked to Alzheimer's disease, with the final candidate binding about 3.6 times better than the patented starting compound. Alibaba also shows the model building playable games in Unity, Unreal Engine, and Godot from a single prompt, including the Unreal Engine sample "Meow of Magic," and creating reusable 3D models in Blender.
  • Safety: On Last Secure Code Bench, the share of functional and secure code rose from 22 to 31.8 percent, and the share secure at all three testing levels nearly doubled, from 8 to 14.7 percent. Alibaba says internal reviewers examined the model's full chains of thought and found no real CBRN capacity for innovation that would give a user a significant capability boost.

Why it matters

Qwen4-Max illustrates a broader industry shift: competing less on raw benchmark peaks and more on cost efficiency per task. On the Artificial Analysis Intelligence Index, it scores 43 at its highest reasoning setting—nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro, the top Chinese model at 46—but costs roughly a third of its predecessor per task. That gap between capability and price is the release's core pitch.

The architecture itself signals where frontier models are heading. Hybrid attention designs that compress context before selective re-examination mirror DeepSeek's sparse-attention approach in V4-Pro, suggesting long-context efficiency is becoming a shared design pattern among Chinese labs rather than a one-off. The largest gains over the predecessor came on HLE-Diamond (up 21 points to 45 percent), AA-Omniscience (up 18), and AutomationBench-AA (up 17).

The Live Delivery feature and the autonomous lab and game-development demos also reflect the industry's agent focus: models that run long tasks on their own need mechanisms for human intervention, and Alibaba is bundling that mechanism with the model. The contrast with Anthropic—which recently barred Claude Code from recreating existing games—shows two labs taking different stances on AI-driven game development.

Who it affects

  • Developers and teams building agents: Lower per-task costs, a one-million-token context window, and the open-sourced Live Delivery plugin directly affect the economics of running long-horizon agent workflows.
  • Enterprise users in industry and finance: Alibaba targets professional workloads explicitly, citing industrial programming and finance deployments with verifiable, traceable outputs.
  • AI researchers and labs: The hybrid attention architecture and the autonomous chemistry demonstrations point to where model design and applied AI in physical labs are converging.
  • Cost-conscious buyers of Chinese frontier models: Qwen4-Max sits second among Chinese models on the Artificial Analysis index but undercuts Kimi K3.1 Pro on price; Tencent Hy4 remains much cheaper ($1.03 per task) but scores eight points lower.
Original article

Qwen4-Max is Alibaba's answer to Opus 5.5, GPT-6, and Kimi K3.1 Pro. The new flagship pairs a more efficient architecture with much lower prices and targets professional work in industry, science, and finance.

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.

A hybrid architecture makes long contexts cheaper to process

Qwen4-Max debuts the revamped architecture of the Qwen4 series. Instead of relying only on standard attention layers, Alibaba combines two methods. Three layers compress the prior context into a condensed state, and a fourth layer then picks out only the parts that matter. That sharply cuts compute costs on long inputs, since the model no longer has to give every part of the context equal attention.

Architecture diagram of Qwen4-Max showing three Gated DeltaNet layers for every Qwen Sparse Attention layer, gated residual read and write steps, MoE blocks, and an N-gram embedding layer at layer 2.

Qwen4-Max compresses prior text across three layers before a fourth layer takes a closer look at only the important sections. | Image: Alibaba

The architecture also adds four parallel connections between model layers, with dynamic gates meant to stabilize the flow of information. An N-gram embedding layer with 328 billion parameters supplies common token sequences as ready-made building blocks. According to Alibaba, training the 2.4-trillion-parameter configuration ran stably. Training loss fell steadily throughout the run with no sudden spikes.

Line chart of Qwen4-Max's training loss across the full pretraining run, falling from 1.8 to just above 1.0.

Training loss declined steadily throughout the run with no loss spikes, according to Alibaba. | Image: Alibaba

Alibaba's internal evaluations measure model quality in bits per byte across knowledge, code, question answering, and STEM. Qwen4-Max-Base scored best among the five checkpoints tested. It improved 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.

Four bar charts compare bits per byte for Qwen4-Max-Base, Qwen3.8-Max-Base, GLM-5.3, DeepSeek-V4-Pro-Base, and Kimi K3.1 Pro across knowledge, code, question answering, and STEM.

Qwen4-Max-Base posts the lowest, and therefore best, values in all four categories. | Image: Alibaba

Lower token use cuts the cost of each task by more than half

The efficiency gains show up in operating costs. Work360 is Alibaba's own benchmark for industrial work tasks across 721 specialized fields. On it, Qwen4-Max uses roughly a quarter to a third fewer tokens than its predecessor to solve the same tasks. On the agent benchmark Terminal-Bench 4.0, token use drops by nearly half.

According to Alibaba, the average cost per Work360 task falls 62 percent, from $4.43 to $1.67. The benchmark score climbs from 47.74 to 52.69 at the same time.

Two charts plot score against tokens used per task on Work360 and Terminal-Bench 4.0 for Qwen4-Max and Qwen3.8-Max at the Low, High, and Xhigh reasoning levels.

Qwen4-Max beats its predecessor on both benchmarks while using far fewer tokens. | Image: Alibaba

Results on external benchmarks are mixed. On long-horizon agent tasks like Agents' Last Exam v1.1, Qwen4-Max scores 46.7 points, the best result 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.

Bar charts compare Qwen4-Max with Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, and Kimi K3.1 across 16 benchmarks, including Terminal-Bench 4.0, Agents' Last Exam v1.1, HealthBench Professional, and Humanity's Last Exam.

Alibaba says Qwen4-Max leads on agent and work tasks, while Opus 5.5 and the GPT models come out ahead on Humanity's Last Exam. | Image: Alibaba

Qwen4-Max ranks second among Chinese models at a third of its predecessor's cost

At its highest reasoning setting, Qwen4-Max scores 43 points on the current Artificial Analysis Intelligence Index. That's nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro, the top Chinese model at 46 points. The biggest gains over the predecessor come on HLE-Diamond (up 21 points to 45 percent), AA-Omniscience (up 18), and AutomationBench-AA (up 17).

This image combines two charts from Artificial Analysis. A bar chart of Intelligence Index v5.0.1 shows Claude Fable 5.2 on top with 65 points and Qwen4-Max at 43 points, with an arrow marking the jump from Qwen3.8-Max at 34 points. A scatter plot compares index scores with cost per task and places Qwen4-Max at about $3, well to the left of Qwen3.8-Max at about $10.

Qwen4-Max gains nine points on the Artificial Analysis Intelligence Index and costs only about a third as much per task as its predecessor. | Image: Artificial Analysis

Artificial Analysis puts the cost of Qwen4-Max at $3.27 per Intelligence Index task. That's 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 are 25 percent lower than the predecessor's, cache hits cost 70 percent less, and the model burns through fewer tokens. Qwen4-Max needs about 105,000 output tokens per benchmark task, compared with 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 coding benchmark FrontierCode 2.0, it ties Kimi at 26 percent and trails Tencent Hy4 at 31 percent.

Through its own API, Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens. Cache hits cost $0.075.

Live Delivery hands control of long-running agents back to users

Like much of the AI industry, Alibaba builds this release around agents that handle long tasks on their own. If users only see results at the end, they have almost no way to fix mistakes. Alibaba's answer is "Live Delivery." The model delivers a usable first version early and builds it out in later passes. Users can step in and change course at any point. Alibaba has open-sourced the accompanying Qwen-Live-Delivery-Plugin.

Alibaba cites several partner 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. It delivers the results as verifiable Excel and PowerPoint files, and every number traces back to its original source.

Alibaba also uses the model internally to run its own cloud infrastructure more efficiently. According to the company, Qwen4-Max worked with the in-house solver MindOpt to cut the cost of allocating compute jobs for data annotation by an average of 10.42 percent. When distributing CPU tasks in data centers, it found optimal solutions more than 14 times faster than the comparison method.

Alibaba's models now run lab experiments on their own

Alibaba also shows the Qwen4 generation at work in physical labs. Paired with Mira, a system from startup Deep Principle, the model ran autonomously for 48 hours, according to Alibaba. In that time, it 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.

Qwen4-Max also worked with PharmMaster, a platform from Beijing-based AI drug discovery startup DP Technology. The target was a compound aimed at an inflammatory response linked to Alzheimer's disease. Over several rounds of literature review, design, and computation, the model optimized 24 candidate molecules. According to Alibaba, 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.

Qwen4-Max aims to build playable games from a single prompt

Alibaba also shows Qwen4-Max working with established game engines like Unity, Unreal Engine, and Godot. From a single natural-language description, the model is supposed to handle the entire development process. That covers scene setup, game logic, and matching external assets, along with materials, UI, and visual effects, until the game is playable.

Scene from the game "Meow of Magic" showing a cat in a wizard's robe and hat inside a medieval hall with stained-glass windows.

Qwen4-Max built a wizard adventure in Unreal Engine in which players, as a cat, battle waves of enemies and bosses. | Image: Alibaba

During development, the model runs the game itself and uses visual and interactive feedback to tweak camera work, character animations, hit reactions, and lighting. Alibaba's sample project is "Meow of Magic," an Unreal Engine adventure about a cat wizard in which players cast spells and fight waves of enemies. According to Alibaba, Qwen4-Max also builds 3D models in Blender that can be reused for games and 3D printing.

Anthropic recently moved its coding assistant Claude Code in the opposite direction. After a wave of unofficial Halo and GTA: Vice City clones, it barred the tool from recreating existing games.

Alibaba says the model writes safer code and offers no real CBRN uplift

On the safety benchmark Last Secure Code Bench, the share of code that is both functional and secure rose from 22 to 31.8 percent, according to Alibaba. The share rated secure at all three testing levels nearly doubled, from 8 to 14.7 percent. Alibaba also tested for CBRN risks (chemical, biological, radiological, and nuclear). Internal reviewers went through the model's full chains of thought, the company says. They 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.

In August, Alibaba released Qwen3.8-Max, its first Max model built for autonomous tasks that run for days. It also has 2.4 trillion parameters, but 95 billion are active per token. At the time, Artificial Analysis found that the model needed 64 steps per task instead of 14 and cost more than twice as much per task as Qwen3.7-Max.

In late August, Alibaba previewed the Qwen4 architecture with Qwen3.8-Flash-Next, which already used an N-gram embedding layer with 51 billion parameters. The company had released the Qwen3.8-Max weights in mid-August. For Qwen4, only the smaller models will get open weights for now.

Moonshot AI followed in October with Kimi K3.1 Pro. At 2.8 trillion parameters, it's even larger, and Moonshot says it sits just behind Claude Opus 5.5 on Terminal-Bench 4.0. Like Alibaba, DeepSeek uses a hybrid attention design with sparse attention in V4-Pro to process long contexts more cheaply. Zhipu AI took a different route and got GLM-5.3's gains entirely from extra post-training.

The original belongs to The Decoder and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.

Other coverage 7