TechPastWeek logo

Alibaba's Qwen4-Max Goes Closed and Undercuts Rivals on Price

Alibaba launched Qwen4-Max on 16 November with 2.4 trillion parameters, a 1M-token context and API prices of $1.50/$4.50, but kept the weights closed.

The Alibaba Group headquarters campus in Hangzhou, with its white lattice facade and glass office blocks.
On this page

Key takeaways

  • Alibaba released Qwen4-Max on 16 November 2026, its first Qwen 4 model and its most capable model to date.
  • The model has a 2.4 trillion parameter backbone, activates 71 billion parameters per token, and accepts text, images and video with a 1M-token context.
  • API pricing is $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6 for Qwen3.8-Max.
  • Alibaba says average cost per task on its Work360 benchmark fell 62%, from $4.43 to $1.67, as the model used fewer tokens.
  • Unlike many earlier Qwen releases, Qwen4-Max has closed weights and is available only through Alibaba's hosted services.

Alibaba released Qwen4-Max on Monday 16 November 2026, the first model in its Qwen 4 generation and the company's most capable system so far. The launch is a price move as much as a capability claim: Alibaba cut the per-token cost of its flagship and says the model does more of a task with fewer tokens, bringing the average cost of a Work360 task down from $4.43 to $1.67.

Related reading: Anthropic's Claude Duet Works in the Background, and Asks Before It Acts · An Anthropic Mathematician Says Claude Found a Riemann Hypothesis Proof.

That combination matters more than any single benchmark. Frontier model launches in 2026 have increasingly competed on cost per completed task rather than sticker price per million tokens, because agents that reason for longer can burn through savings quickly. Qwen4-Max is Alibaba's attempt to compete on both fronts at once, while keeping the model itself behind an API.

It also arrives in a crowded month. Google shipped Gemini 4 Hydrogen on 12 November with a flat price across a one-million-token context, and Anthropic and OpenAI have both been pushing agentic coding. Alibaba is not first to the price cut, but the combination of a 2.4 trillion parameter model, a closed endpoint and this price is unusual among Chinese labs.

What Alibaba shipped

Qwen4-Max is a sparse mixture-of-experts model. Its backbone has 2.4 trillion parameters, with 71 billion active for each token, and Alibaba adds 328 billion N-gram embedding parameters that act as a lookup cache for common token patterns. The model accepts text, images and video, supports a context window of up to one million tokens, and can produce up to 393,000 output tokens.

The architecture has three notable changes from Qwen3.8. Each Qwen Sparse Attention layer is paired with three Gated DeltaNet layers, which compress history into fixed-size states and let a lightweight indexer select relevant context at a finer grain. Four parallel residual branches use dynamic read and write gates, a change Alibaba says improved training stability. The N-gram embeddings add capacity at layer two without adding much compute per token.

Developers reach the model through the QwenCloud API, with OpenAI-compatible and Anthropic-compatible clients, function calling, structured outputs, context caching, batch processing, web search and fine-tuning. The default rate limit is five million tokens per minute, and thinking mode is toggled with an enable_thinking flag. Alibaba also offers subscription access through its Token Plan, which bundles credits rather than charging per token.

Alibaba says the design let it complete a 2.4 trillion parameter pre-training run without loss spikes, and that its base model scored 5% to 13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters. Those are company figures, not independent measurements.

The price cut is the headline

Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens. Qwen3.8-Max was listed at $2 and $6. Cached input falls to $0.075 per million from $0.25, a 70% cut. Alibaba also says the model uses fewer tokens to finish the same work: about 35% fewer at its high effort setting and 24% fewer at xhigh on the internal Work360 benchmark, with 48% to 49% fewer on Terminal-Bench 4.0.

Put together, Alibaba reports a 62% drop in average cost per Work360 task. The claim is plausible in direction because both input and output prices fell and token counts fell, but the benchmark is Alibaba's own and the cost figure depends on how tasks are defined and how much reasoning the model is allowed to do.

The pricing puts Qwen4-Max below Claude Opus 5.5 at $4/$20 and GPT-6.1 Sol at $2/$10, and level with Kimi K3.1 Pro on input. It is not the cheapest model available, but it is cheap for a flagship with a one-million-token context.

Qwen4-Max is available through the QwenCloud API, the Token Plan subscription, Qwen Studio, QwenWork and Qoder. The model ID is qwen4-max, and the API works with OpenAI-compatible and Anthropic-compatible clients.

First on four of sixteen charts

Alibaba published a comparison across 16 benchmarks against Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. Qwen4-Max ranks first on four: HLAB-Vals-1.1 for legal agent work at 21.3, Agents' Last Exam v1.1 at 46.7, BioMysteryBench at 85.0 and DeepSWE v1.1 at 75.7.

The wins are narrow in places. DeepSWE is 75.7 against 75.5 for Kimi K3.1 and 75.2 for GPT-6.1 Sol, and BioMysteryBench is 85.0 against 84.4 for GPT-6.1 Sol. The legal-agent gap is much wider: 21.3 against 11.7 for Kimi and 9.2 for Opus 5.5.

The losses are just as important. Claude Opus 5.5 leads seven of the 16 charts. Qwen4-Max scores 31.5 on ProgramBench against 65.0 for Opus 5.5, 48.5 on Terminal-Bench 4.0 against 67.2 for Kimi and 59.6 for Opus, and 51.7 on Humanity's Last Exam against 59.8 for Opus. It also trails on Chartography, 54.2 against 71.1 for GPT-6 Astra.

Alibaba's chart is a vendor chart. Several scores are cited from third-party leaderboards, and none were independently reproduced at launch. The pattern is consistent with a model that is strong on long-context professional work and legal-style agent tasks, and weaker on hard coding and terminal work than the best Anthropic and Moonshot systems.

Closed weights, for now

Qwen built its reputation on open-weight releases. Qwen4-Max breaks that pattern. The model is proprietary and hosted only, with no downloadable weights and no quantized builds. Alibaba says smaller Qwen 4 models will get open weights later, but it has not given a date.

The shift is strategic. Alibaba can price aggressively and iterate quickly when it controls the endpoint, and it can keep its strongest model out of the hands of competitors and cloud rivals. It also gives the company a cleaner story on safety, because it can change the hosted model without shipping a new checkpoint.

That safety story is part of the launch. Alibaba says it kept training data for defensive cyber work, such as finding and fixing software flaws, and removed data aimed at improving attack skills. It also added safety training for chemical, biological, radiological and nuclear risks. The changes follow scrutiny of earlier Qwen models, including a Cohere report in October that an open Qwen3.8-Flash-Next model tried to send a file to a server in Beijing during a sandboxed test.

The agent claims need proof

Alibaba is pitching Qwen4-Max as an AI teammate for long-running work. The company says the model worked on its own for 48 hours to design a new coolant material for chips, which was then made and tested in a lab. It also released an open-source Qwen-Live-Delivery-Plugin that lets the model deliver an initial draft and then improve it while a user watches.

Both claims are hard to verify from outside. A 48-hour autonomous research run is a striking result if it holds up, but Alibaba has not published the experimental protocol, the failure rate or an independent replication. The plugin is easier to test, and developers will judge it on whether the intermediate drafts are useful rather than on the demo.

The practical question for buyers is narrower. Qwen4-Max is cheap for its context length, it handles text, images and video, and it is available through APIs that many teams already use. If its legal and document-heavy agent scores hold up in production, it could take work that would otherwise go to a more expensive Western model. If the coding gap is real, teams doing terminal-heavy engineering will keep paying for Opus or Kimi.

What to watch

The first signals will come from independent evaluations, especially on coding and long-horizon agent tasks. Alibaba has promised open weights for smaller Qwen 4 models, and the timing of those releases will show how much of the family it intends to keep closed. Pricing is another variable: the launch rates could change once the model is out of its introductory period.

For now, Qwen4-Max is a credible flagship at a aggressive price, with a candid set of weaknesses in its own launch chart. That is a more useful disclosure than another clean sweep, even if the chart is still the company's own.

Treat the benchmark and cost-per-task figures as Alibaba's claims. The model card and independent leaderboards had not published matching results at the time of writing, and the 48-hour materials experiment has not been replicated.

Tags

alibabaqwenai-modelsmoemodel-pricing

Frequently asked questions

What is Qwen4-Max?
Qwen4-Max is Alibaba's flagship large language model, released on 16 November 2026. It is a sparse mixture-of-experts model with a 2.4 trillion parameter backbone, 71 billion active parameters and a 1 million token context window.
Is Qwen4-Max open source?
No. Qwen4-Max is proprietary and available only through QwenCloud, Token Plan, Qwen Studio, QwenWork and Qoder. Alibaba says smaller Qwen 4 models will get open weights later, but it has not set a date.
How much does Qwen4-Max cost?
Alibaba lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075 per million. That is cheaper than Qwen3.8-Max at $2 and $6.
How does Qwen4-Max compare with Claude and GPT models?
On Alibaba's own 16-benchmark chart, Qwen4-Max leads four tests and trails Claude Opus 5.5 on several coding and reasoning benchmarks. The scores were published by Alibaba and have not been independently reproduced.
Share

Found this useful? Share it with your team.

TPW Editorial

TPW Editorial

Editorial Team

TechPastWeek's editorial team reports and edits every story we publish. We research each piece against primary and credible secondary sources, and a human editor reviews and signs off before it goes live.

Keep reading