Technology

Alibaba unveils Qwen4-Max AI model with 1 million token context

Illustration: Tezkun / AI. Not a photograph of the event.

Alibaba has released its new flagship artificial intelligence model, Qwen4-Max. According to the company, it consumes significantly fewer tokens than its predecessor Qwen3.8-Max and supports image processing.

The model has a total of 2.4 trillion parameters—the same as Qwen3.8-Max—yet only 71 billion of them are activated per token. Qwen4-Max is built on a mixture-of-experts architecture. It accepts up to 991,000 input tokens, while a single response can span up to 131,000 tokens, reaching up to 262,000 in reasoning mode.

According to Alibaba, the newcomer expends fewer tokens on the same tasks: roughly 48–49% less on Terminal-Bench 4.0 and 24–35% less in the internal Work360 benchmark, with the variance depending on the reasoning mode. The average cost per task in Work360 dropped from $4.43 to $1.67, representing a 62% decrease (or 60%, according to 3DNews). In QwenCloud, one million input tokens cost $1.5, one million output tokens cost $4.5, and cached input costs $0.075.

In the preliminary Code Arena WebDev leaderboard, the model debuted in sixth place with 1,750 points, while placing 14th in Agent Arena; in both cases, this represents the highest score among Chinese models. Alibaba's internal benchmark results have not yet been independently verified. The company also stated that it removed data suitable for "offensive" cybersecurity operations from the training set, preserving defensive capabilities such as vulnerability scanning.

Qwen4-Max is already available via the QwenCloud API, as well as in Qwen Studio, QwenWork, Qoder, and OpenCode Go. The model itself remains closed-source, but Alibaba pledges to release smaller versions of Qwen4 with open weights; no date has been announced yet.

What it means

Qwen4-Max is already available via the QwenCloud API and across Qwen Studio, QwenWork, Qoder, and OpenCode Go services. One million input tokens cost $1.5, output tokens cost $4.5, and cached input costs $0.075 per million. The weights for Max itself will not be opened; junior Qwen4 models are promised with open weights, with no date set.

Published 16 November 2026, 23:02 · No updates · Russian original

Discussion
  1. Loading comments…
No links or insults. One comment per 20 seconds.

Comments on this story live under its post in our Telegram channel. Comment on Telegram