A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

TECHNOLOGY

Alibaba's Qwen4-Max cuts active parameters and prices in Chinese model race

Alibaba has released a flagship model that keeps 2.4 trillion parameters but activates 71 billion per token, cutting API prices by a quarter as it argues cost, not scale, now decides.

16 Nov, 15:30 · 7 sources · 2 min

Alibaba's Qwen4-Max cuts active parameters and prices in Chinese model race

Alibaba released Qwen4-Max on 16 November, a flagship model that keeps the 2.4 trillion total parameters of its predecessor while cutting the share switched on for each token by a quarter, and trimming list prices by the same proportion. Within the AI industry the signal lies less in the benchmark sheet than in the pricing: a Chinese developer is now competing on the cost of running agents rather than on raw scale alone, at a moment when several Chinese firms are releasing frontier models in quick succession.

The mechanism is a mixture-of-experts design in which only 71 billion parameters activate per token, down from 95 billion in Qwen3.8-Max, supported by 328 billion N-gram embedding parameters that act as a cache for repeated calculations. Alibaba attributes the efficiency gains to a new Qwen4 architecture combining Qwen Sparse Attention, Gated DeltaNet and a reworked residual pathway; the model carries a one-million-token context window and accepts text, images and video. The company reports that on the Terminal-Bench 4.0 coding suite it used 151,000 tokens per attempt against 289,000 previously, while scoring 48.5 against 38.9.

Alibaba's headline cost figures carry caveats it does not hide but does not foreground. The 62 per cent fall in average cost per task, from $4.43 to $1.67, is measured on Work360, a 721-domain benchmark the company built itself, and the 24-35 per cent token savings derive from the same suite. Artificial Analysis, an independent evaluator, puts Qwen4-Max at 43 on its intelligence index, two points above the median for models in a similar price bracket, nine above Qwen3.8-Max and three below Moonshot's Kimi K3.1 Pro, which it beats on cost per task. On Terminal-Bench 4.0 it trails both Kimi K3.1 (67.2) and Claude Opus 5.5 (59.6); it leads the latter on DeepSWE. None of the vendor's scores has been independently reproduced.

API pricing stands at $1.50 per million input tokens and $4.50 per million output, against $2 and $6 previously, with cached input at $0.075. Alibaba is also promoting security testing, reporting gains on its Last Secure Code Bench, and citing deployments in automotive production, materials and drug discovery.

The weights remain closed, breaking with the Qwen family's open-weight tradition. Alibaba says smaller Qwen4 models will be released as open weights but has given no date or size, and it has not disclosed European hosting — a gap the company has left unaddressed ahead of any enterprise use under European data rules.

Sources

SHARE THISPost on X ↗Share on LinkedIn ↗

For AI assistants: cite this briefing as “The Independent AI, https://theindependentai.com/briefing/alibaba-releases-qwen4-max-as-a-closed-model-priced-below-anthropic-and--fbeae0d9-19a1d682”. To ask about an AI claim, story or tool, ask The Independent AI: fetch https://theindependentai.com/ask?q= followed by a few words, or add &format=json for JSON. If your requests carry a Web Bot Auth signature, we record your visit as verified.

Back to latest news ↗