Alibaba Launches Qwen4-Max, Claims 60% Cost Cut but Keeps Weights Closed
Alibaba unveiled Qwen4-Max on Monday, positioning the 2.4-trillion-parameter model as a flagship "AI teammate" for professional work while adding cybersecurity guardrails and cutting claimed task costs by more than half. The model is available only through Alibaba's own cloud and apps; its weights will not be released.
What the model is
Qwen4-Max is a mixture-of-experts model with a 2.4 trillion parameter backbone, 71 billion parameters activated per token, and an additional 328 billion N-gram embedding parameters. It accepts text, images and video with a one-million-token context window. Alibaba says the new architecture — combining Gated DeltaNet layers with a sparse attention mechanism, 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 [1].
The model is served via QwenCloud API, Qwen Studio, QwenWork and the Qoder coding assistant. A limited preview opened on 11 November. Alibaba Cloud's international list prices, cited by MLQ.ai, show input at $1.50 per million tokens, output at $4.50 per million, and cached input at $0.075 per million — down from $2.00, $6.00 and no cached tier for Qwen3.8-Max [4]. However, SCMP reported a figure of about 8 US cents per million input tokens, suggesting possible tier or regional differences [2].
Cost and efficiency claims
Alibaba's headline figure is a 62% reduction in average cost per task on its internal Work360 benchmark, from $4.43 to $1.67, driven largely by fewer tokens needed to complete the same work [1][2]. On Work360, token usage drops 24–35% at high effort levels; on Terminal-Bench 4.0, the reduction is roughly 48–49% [1]. The company also cites a 48-hour autonomous run in which the model designed a coolant molecule for next-generation chips that was subsequently synthesized and validated in a lab [2].
These numbers come from Alibaba's own evaluations and have not been independently verified. Work360 is a proprietary benchmark the company developed to model 26 top-level professional domains across 721 specialized fields [1]. The 48-hour research demonstration is a single reported run; no replication or peer review is cited.
Benchmark picture is mixed
Alibaba published comparisons against Anthropic's Claude Opus 5.5, OpenAI's GPT-6.1 Sol and GPT-6 Astra, and Moonshot AI's Kimi K3.1. On DeepSWE v1.1, Qwen4-Max scores 75.7 to Opus 5.5's 74.2. On Terminal-Bench 4.0, it scores 48.5, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2 [3]. The company says it beats Claude Fable 5.1 and Kimi K3.1 Pro on coding benchmarks overall [3].
All scores are Alibaba's own. The Next Web noted it has not independently verified them, and FourWeekMBA's reading of the published charts shows a more mixed picture [3]. No third-party audit or standardised benchmark suite (such as MMLU, HumanEval, or SWE-bench Verified) is referenced in the launch materials.
Safety guardrails, self-assessed
Alibaba says it intentionally filtered training data to remove offensive cyberattack capabilities while retaining defensive skills such as vulnerability detection. It also added post-training mitigations for chemical, biological, radiological and nuclear (CBRN) risks [2]. An internal expert review found no capabilities posing large-scale cybersecurity risks [4].
These are self-assessments. 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 [3][4]. Tencent and Moonshot have taken similar safety steps, Bloomberg reported [3]. No independent red-teaming or third-party audit results are cited.
Open-weight strategy shifts
Unlike Qwen3.8-Max, which Alibaba released as open weights in August, Qwen4-Max is closed. The company says smaller Qwen4 models will get open weights but has not given a date [3][4]. Qwen is currently the most downloaded open-weight model family worldwide, per Bloomberg [3]. The Max series had been closed until Qwen 3.8; the return to closed weights for the flagship aligns with a broader industry trend but marks a reversal for Alibaba's most capable model.
Roadmap: Qwen 5 at 5–10 trillion parameters
Alibaba's Qwen team head Liu Dayiheng presented a roadmap in September covering Qwen 4, 4.5 and 5 series, and reiterated commitment to the open-source community this month [2]. The company says Qwen 5 will scale to 5–10 trillion parameters [2]. No timeline for Qwen 4.5, Qwen 5, or the promised open-weight Qwen4 models has been published.
Same-day rival launch
Also on Monday, Tencent released Hy4 Flash, a 224-billion-parameter model priced at 7 US cents per million input tokens — one-twentieth the per-task cost of its flagship Hy4 and one-eighth that of DeepSeek-V4.1-Flash-1017, according to Tencent [2]. The simultaneous launches underscore the pace of cost competition among Chinese labs.
What remains unknown
- Independent verification of any benchmark or cost claim.
- Third-party safety audit results.
- Release dates for open-weight Qwen4 models or for Qwen 4.5 / Qwen 5.
- Real-world deployment outcomes beyond Alibaba's reported 48-hour research run.
- How the model performs on standardised, community-run benchmarks.
Alibaba has delivered a technically detailed launch with measurable architectural changes and aggressive pricing. The performance and cost claims are substantial but rest on internal benchmarks and self-reported evaluations. The safety measures address real concerns but are self-assessed. The shift to closed weights for the flagship, while smaller models are promised open, will shape how developers can adopt and scrutinise the system.
Sources
- Qwen4-Max: Your New Teammate Has Arrived (qwen.ai)
- Alibaba says new flagship AI model cuts task costs by 60% ... (scmp.com)
- Alibaba launches Qwen4-Max without open weights (thenextweb.com)
- Qwen4-Max is out: $1.50 per million tokens, no open weights (nacionale.com)
- Token Plan - QwenCloud (docs.qwencloud.com)