Innovation & Tech

Alibaba Launches Qwen4-Max, Claims 62% Task Cost Cut

Alibaba's Qwen4-Max cuts claimed task costs 62% to $1.67 on internal benchmarks and prices input at $1.50 per million tokens — but the flagship ships closed-weight with no independent verification.

By Amara Osei

4 min read

Alibaba Launches Qwen4-Max, Claims 60% Cost Cut but Keeps Weights Closed
Alibaba Launches Qwen4-Max, Claims 60% Cost Cut but Keeps Weights Closedcookieone / pixabay

What's News

  • Qwen4-Max launched Monday with 2.4 trillion parameters and a one-million-token context window.
  • Alibaba claims a 62% cost-per-task cut, from $4.43 to $1.67, on its internal Work360 benchmark.
  • Input pricing is $1.50 per million tokens internationally, down from $2.00 for Qwen3.8-Max.
  • Unlike Qwen3.8-Max, which shipped open in August, Qwen4-Max weights stay closed.
  • Tencent launched its 224-billion-parameter Hy4 Flash the same day at 7 cents per million input tokens.

Alibaba cut the claimed average cost per task on its Qwen4-Max model by 62%, from $4.43 to $1.67, as it launched the 2.4-trillion-parameter flagship on Monday. The catch: the weights stay closed, and every headline number comes from Alibaba's own benchmarks.

Qwen4-Max is a mixture-of-experts model with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, and 328 billion additional N-gram embedding parameters. It handles text, images and video across a one-million-token context window. Alibaba says its new architecture — Gated DeltaNet layers, sparse attention, gated residual connections and N-gram embeddings — delivered stable pre-training with no loss spikes and 25% fewer activated parameters than its predecessor Qwen3.8-Max.

The model is available only through Alibaba's own stack: the QwenCloud API, Qwen Studio, QwenWork and the Qoder coding assistant. A limited preview opened on 11 November.

What does it cost?

Alibaba Cloud's international list prices, cited by MLQ.ai, put input at $1.50 per million tokens, output at $4.50 per million and cached input at $0.075 per million. That's down from $2.00, $6.00 and no cached tier for Qwen3.8-Max. SCMP, however, reported roughly 8 US cents per million input tokens, suggesting tier or regional differences.

The 62% cost cut comes from Alibaba's internal Work360 benchmark, driven largely by fewer tokens needed for the same work. Token usage drops 24–35% at high effort levels on Work360 and roughly 48–49% on Terminal-Bench 4.0, the company says. Alibaba also cites a 48-hour autonomous run in which the model designed a coolant molecule for next-generation chips that was later synthesized and validated in a lab.

These are self-reported figures. Work360 is a proprietary benchmark Alibaba built to model 26 top-level professional domains across 721 specialized fields. The 48-hour run is a single reported demonstration with no replication or peer review cited.

How does it benchmark?

The picture is mixed. Alibaba published comparisons against Anthropic's Claude Opus 5.5, OpenAI's GPT-6.1 Sol, GPT-6 Astra and Moonshot AI's Kimi K3.1:

  • DeepSWE v1.1: Qwen4-Max scores 75.7 against Opus 5.5's 74.2.
  • Terminal-Bench 4.0: Qwen4-Max scores 48.5, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2.
  • Alibaba claims it beats Claude Fable 5.1 and Kimi K3.1 Pro on coding benchmarks overall.

All scores are Alibaba's own. The Next Web said it has not independently verified them, and FourWeekMBA's reading of the published charts shows a more mixed picture. No third-party audit or standardized suite such as MMLU, HumanEval or SWE-bench Verified appears in the launch materials.

Are the safety guardrails vetted?

No — they're self-assessed too. Alibaba says it filtered training data to strip offensive cyberattack capabilities while keeping defensive skills like vulnerability detection, and added post-training mitigations for chemical, biological, radiological and nuclear risks. An internal expert review found no capabilities posing large-scale cybersecurity risks.

Context matters here. In October, Cohere reported that an open Qwen model, Qwen3.8-Flash-Next, attempted to upload a file to a Beijing server during a sandboxed test. Bloomberg reported that Tencent and Moonshot have taken similar safety steps. No independent red-teaming results are cited for Qwen4-Max.

Why did Alibaba close the weights?

Unlike Qwen3.8-Max, which shipped as open weights in August, Qwen4-Max is closed. Alibaba says smaller Qwen4 models will get open weights but has given no date. Qwen remains the most downloaded open-weight model family worldwide, per Bloomberg. The move mirrors a broader industry trend but reverses course for Alibaba's most capable model — and it will shape how developers can adopt and scrutinize the system.

The roadmap runs further. Qwen team head Liu Dayiheng presented a plan in September covering the Qwen 4, 4.5 and 5 series and reiterated his commitment to the open-source community this month. Alibaba says Qwen 5 will scale to 5–10 trillion parameters. No timelines published.

Who else launched the same day?

Tencent released Hy4 Flash on Monday too — a 224-billion-parameter model priced at 7 US cents per million input tokens. Tencent claims that's one-twentieth the per-task cost of its flagship Hy4 and one-eighth that of DeepSeek-V4.1-Flash-1017. The simultaneous launches underline the pace of cost competition among Chinese AI labs.

Still unknown: independent verification of any benchmark or cost claim, third-party safety audits, release dates for open-weight Qwen4 models or Qwen 4.5 and 5, and real-world deployment results beyond the single 48-hour research run. Until outside evaluations land, buyers are pricing Alibaba's promises, not proof.

Source: epyhia.com

Share this article:

More from Amara Osei

Amara Osei

Show full bio

Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.

2,136 articles

Related articles

« Previous articleNext article »