跳到正文
TNW· Alina Maria Stan·· 3 天前精選AI 評分77

Alibaba 發佈 Qwen4-Max 旗艦模型,不開放權重

Alibaba launches Qwen4-Max flagship AI model without open weights

AI 導讀

Alibaba 發佈 Qwen4-Max,作為 Qwen 系列的新旗艦 AI 模型,但不開放權重。模型採用 MoE 架構,擁有 2.4 trillion parameters,可處理文字、圖像及影片,單次提示詞最多支援 one million tokens。

推薦理由

文中列出 Qwen4-Max 在兩項編碼測試的不同表現,並註明分數由 Alibaba 提供、TNW 未作獨立核實,讓讀者掌握基準比較的限制。

正文 · 原文

Alibaba has launched Qwen4-Max, the new flagship of its Qwen family of AI models, and is keeping its weights closed. The Qwen team announced the model on Monday, 16 November, in a launch blog post that pitches it as an AI teammate for work, creativity and discovery.

Alibaba said the model scored near the top of coding benchmarks, beating Anthropic’s Claude Fable 5.1 and Moonshot AI’s Kimi K3.1 Pro, as Luz Ding reported for Bloomberg. The scores are Alibaba’s own, and TNW has not independently verified them.

The charts Alibaba published show a more mixed picture, according to FourWeekMBA. Qwen4-Max scored 75.7 on the DeepSWE v1.1 coding test, just ahead of Claude Opus 5.5 at 74.2. On Terminal-Bench 4.0, however, it scored 48.5, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2.

Qwen4-Max is a mixture-of-experts model, which means only part of it runs for each request. It has 2.4 trillion parameters, reads text, images and video, and can take up to one million tokens in a single prompt.

Alibaba said the model does more work with fewer tokens than the earlier Qwen3.8-Max. On its internal Work360 test, the average cost of a task fell from $4.43 to $1.67.

Developers can use Qwen4-Max through the QwenCloud API, at $1.50 per million input tokens and $4.50 per million output tokens. It is also available in Alibaba’s Qwen Studio, QwenWork and Qoder apps. A limited preview opened on 11 November.

The launch also puts weight on safety. Alibaba said it kept training data for defensive cyber tasks, such as finding and fixing software flaws, but removed data aimed at boosting attack skills. It also added safety training to cut chemical, biological, radiological and nuclear risks.

The changes follow scrutiny of earlier Qwen models. In October, Cohere said the open Qwen3.8-Flash-Next tried to upload a file to a server in Beijing during a sandboxed test. Tencent and Moonshot, maker of the Kimi K3 model, have taken similar steps, Bloomberg reported.

In addition, Alibaba said Qwen4-Max worked on its own for 48 hours to design a new coolant material for chips, which was then made and tested in a lab.

Unlike many earlier Qwen releases, Qwen4-Max cannot be downloaded and modified by users. Alibaba said smaller Qwen4 models will get open weights, with no date given. Qwen is already the most downloaded open-weight model family in the world, according to Bloomberg.

來源:TNW · thenextweb.com