N環球 N GLOBE
China Technology

Alibaba launches flagship Qwen4-Max — and keeps its weights closed for the first time

· N GLOBE

Alibaba on 16 November officially released Qwen4-Max, its most capable model yet: 2.4 trillion parameters and a one-million-token context window, with coding scores it claims lead rivals — but the flagship's weights stay closed, a first for Qwen.

Alibaba on Monday officially released Qwen4-Max, its new flagship artificial intelligence model. The most capable Qwen model so far, it uses a mixture-of-experts architecture with 2.4 trillion total parameters and a one-million-token context window, built for coding, office work and long-running AI agents — but unlike previous Qwen flagships, Alibaba is not publishing the model's weights, offering it only through its cloud platform.

What was released

Alibaba announced the release on 16 November (Beijing time), after opening a two-week limited preview on 11 November. The company says Qwen4-Max is a mixture-of-experts (MoE) model with 2.4 trillion total parameters, activating 71 billion per token; it handles up to one million tokens of context and up to about 393,000 output tokens, reads text, images and video, and comes with built-in tools such as web search and a code interpreter. It is compatible with OpenAI and Anthropic API formats and is available through Alibaba's QwenCloud platform, including in Qwen Chat and the Qwen App.

Mixed benchmark results

Alibaba's own benchmark figures, which have not been independently verified, are mixed. On the DeepSWE v1.1 software-engineering test, Qwen4-Max scored 75.7, just ahead of Anthropic's Claude Opus 5.5 at 74.2; but on Terminal-Bench 4.0 it scored 48.5, behind Opus 5.5 at 59.6 and Moonshot's Kimi K3.1 at 67.2. The Artificial Analysis Intelligence Index gives Qwen4-Max 43 points, nine more than its predecessor Qwen3.8-Max and three behind Kimi K3.1 Pro, at about 11% lower cost per task. Alibaba also says the model beats Claude Fable 5.1 and Kimi K3.1 Pro on coding benchmarks, and its internal Work360 test shows average cost per task falling from $4.43 to $1.67 against Qwen3.8-Max. In arena rankings, Qwen4-Max placed sixth in the Code Arena WebDev ranking with 1,750 points, 74 ahead of Qwen3.8-Max, making Alibaba the top-ranked Chinese lab and fifth overall.

First flagship without open weights

The most-watched change is the licensing. Previous Qwen flagships were released with open weights for anyone to download and run; Qwen4-Max is closed and API-only, the first time a Qwen flagship has not been opened. Alibaba says smaller Qwen4 models will be released with open weights, with no date given; Qwen4-27B, one of them, can run on a device with 16GB of memory. Qwen remains the world's most downloaded open-weight model family, and reporting says no government order was behind the decision to withhold the flagship's weights.

Tencent launches on the same day

On the same day, Tencent launched Hy4 flash, a lower-cost model with a one-million-token context window, priced at $0.07 per million input tokens and $0.281 per million output tokens, following the Hy4 release on 4 November. The two launches on one day show competition among Chinese model makers shifting from raw capability towards cost and real-world deployment.

What comes next

The market will now watch for two things: when the smaller Qwen4 models arrive with open weights, and whether independent evaluations confirm Alibaba's benchmark claims. The Qwen4-Max preview goes offline on 25 November, with the full release now serving users.

💬 Comments

No comments yet. Be the first!