Models

Alibaba Closes Qwen4-Max Weights, Sells Flagship at $1.50 per Million Tokens

Alibaba's 2.4-trillion-parameter Qwen4-Max ships closed-weights only, priced at $1.50 per million input tokens — a reversal from August's open Qwen3.8-Max release.

Alibaba Keeps Qwen4-Max Closed and Sells Access at $1.50 per Million Input Tokens
Alibaba Keeps Qwen4-Max Closed and Sells Access at $1.50 per Million Input Tokenscookieone / pixabay
By James Calloway6 min read

Why it matters

  • Qwen4-Max, a 2.4-trillion-parameter MoE model, released November 15-16, 2026 with closed weights, hosted only via QwenCloud API.
  • Pricing: $1.50 per million input tokens and $4.50 per million output tokens in Singapore; Beijing/global regions run $1.413/$4.24.
  • Work360 benchmark: average cost per task fell from $4.43 to $1.67; token use dropped 35% at High reasoning effort.
  • Arena debuted Qwen4-Max at No. 14 overall on Agent Arena, the top Chinese model; Vals AI ranked it 11th.
  • It trails rivals on Terminal-Bench 4.0 (48.5 vs 67.2 for Kimi K3.1) and ProgramBench (31.5 vs 65.0 for Claude Opus 5.5).

Alibaba released Qwen4-Max, a 2.4-trillion-parameter flagship model, late Sunday U.S. time — and for the first time since August, its strongest model is one developers cannot download. The Qwen team is offering the model only as a hosted service through the QwenCloud API and apps including Qwen Studio, QwenWork and Qoder. Smaller Qwen4 models will get open weights later, the team said, without committing to a date.

That is a reversal from the last generation. The core of Qwen3.8-Max shipped with open weights in August as Qwen3.8-2.4T-A95B, a text-only release and the first Max-class Qwen to be open. Qwen is already the most downloaded open-weight model family in the world, according to Bloomberg, as cited by TNW. Alibaba is now betting it can sell the closed flagship on price instead.

The pricing is aggressive. QwenCloud lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens in its Singapore region, with cache reads at 7.5 cents per million. Beijing and global deployments run slightly lower, at $1.413 and $4.24, and thinking and non-thinking modes cost the same. For comparison, OpenAI lists GPT-6.1 Sol at $2 and $10 and its mid-tier GPT-6.1 Terra at $1 and $5.

The stakes are straightforward: Alibaba built its AI reputation on openness, and Qwen4-Max tests whether that reputation survives a paywalled flagship.

What changes under the hood?

Qwen4-Max is a mixture-of-experts model (a design where only a subset of the network activates for each token) that activates 71 billion of its 2.4 trillion backbone parameters per token. Alibaba Cloud said that is 25% fewer active parameters than Qwen3.8-Max at the same scale. The model reads images and video natively and accepts up to 1 million tokens of context.

There are limits developers should note. Thinking plus the final answer is capped at 262,144 tokens, and the developer docs state that fine-tuning and batch inference are not supported. Those omissions matter for enterprise buyers who want customization — Alibaba did not say whether either will arrive later.

The launch post claims the model does more with fewer tokens than its predecessor. On Work360, Alibaba's in-house benchmark of industry tasks across 721 specialized fields, token use fell about 35% at High reasoning effort and 24% at Xhigh. Average cost per task dropped from $4.43 to $1.67. On Terminal-Bench 4.0, token use fell by 48% to 49%. The efficiency claims come from Alibaba's own measurements, not independent verification.

The Qwen team announced the release on X: "Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!" The post highlighted the new MoE architecture, native vision and the 1M-token context window.

How does it score against rivals?

Alibaba's own comparison chart is mixed, and the company published the weak numbers alongside the strong ones. Qwen4-Max scored 75.7 on the DeepSWE coding benchmark, narrowly ahead of Claude Opus 5.5 at 74.2 and Kimi K3.1 at 75.5. It posted 21.3 on the HLAB-Vals legal agent test, where no rival in the chart topped 11.7.

It trailed elsewhere. On Terminal-Bench 4.0 it scored 48.5, against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1. On ProgramBench, a black-box software replication test, it managed 31.5 to Opus 5.5's 65.0 — a gap large enough that Alibaba's headline claims about coding strength need qualification.

Independent leaderboards place it as the strongest Chinese model, behind several U.S. rivals. Arena said Qwen4-Max debuted at No. 14 overall on its Agent Arena, the top Chinese model there, and ranked Alibaba the No. 5 lab overall, behind Anthropic, Google, OpenAI and Meta. On Code Arena's WebDev board it landed at No. 6 with a preliminary 1,750 points. Vals AI ranked it 11th on the Vals Index, also the highest of any Chinese model, with notable gains over Qwen3.8-Max on coding and finance tasks.

The pattern: best-in-class from China, but not best-in-class overall on most third-party boards.

Why is Alibaba pitching 48-hour agent runs?

Alibaba is positioning Qwen4-Max for agents that work around the clock, and its demonstrations aim at that claim. In one, the model ran for 48 hours on Deep Principle's Mira lab platform, built a candidate space of 2.7 million molecules and directed an automated wet lab through four design-and-test loops before producing an experimentally validated chip coolant molecule, the Qwen team wrote.

In an automotive production test at SANF and Pacific Union, the manual effort to program a single workpiece fell from about 475 minutes to 24. The team also open-sourced a Live Delivery plugin that has an agent hand over a usable first version early and revise it as users step in. None of these results come with independent replication.

What did Alibaba do about cybersecurity risk?

The company said it filtered cyber training data by task intent, keeping material for vulnerability detection and remediation and removing data meant to strengthen attacks. A review by Alibaba's internal security experts found that "no capabilities posing large-scale cybersecurity risks have been observed," the launch post says. Note the reviewer: internal, not external.

Moonshot AI took a different route last month, removing cyber-related data from training entirely ahead of Kimi K3.1 Pro's open-weight release. Alibaba's intent-based filtering is a finer-grained approach, but it depends on the classifier working as described.

On secure code specifically, Alibaba reported that Qwen4-Max's pass rate for code that is both functional and secure on Last Secure Code Bench rose to 31.8%, from 22.0% for Qwen3.8-Max.

Where can developers get it?

The model is not locked to Alibaba's cloud. Qwen said it is live on OpenRouter and Vercel's AI Gateway, and QwenCloud serves it from Singapore, Beijing, Frankfurt, Hong Kong, Tokyo and Virginia.

The open question is timing. Alibaba has not said when the open-weight Qwen4 models will arrive, and until they do, the company's open-source credibility rests on the previous generation.

The takeaway

Alibaba is charging $1.50 per million input tokens for a closed flagship that beats every other Chinese model on third-party boards but trails Anthropic and Moonshot on several agentic coding tests — and the open-weight release that made Qwen the world's most downloaded model family now has no date attached.

Source: webpronews.com

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Alibaba Cloud Ships Qwen4-Max With 25% Lower API Prices in Singapore
  2. Alibaba's Qwen4-Max undercuts rivals at $3.27 per benchmark task
  3. Thinking Machines' Inkling 2 Small takes on China's open-weight lead
  4. Saluki 27B: A 7.89 GB, 2-Bit Qwen3.8 That Beats the 54 GB Original at Tool Calling
  5. Meta Releases Muse Spark Open: 816B Open-Weight MoE Runs on Two Mac Studios

« Previous articleNext article »