Alibaba has unveiled Qwen4-Max, the first model of its Qwen4 generation. Offered as a proprietary, hosted-only service, it accepts text, images, and video, and comes with a reduction in API pricing and a claimed 62% decrease in average cost per task.
A Mixture of Experts Architecture
Qwen4-Max is a multimodal MoE model (mixture of experts), an architecture where only a fraction of the parameters is activated per token, i.e., per processed text fragment. The model is built on a foundation of 2.4T parameters, with 71B active per token, plus 328B N-gram Embedding parameters. Its context window reaches 1M tokens, with a maximum output of 393K tokens.
Alibaba describes three evolutions compared to the previous generation. The first combines a QSA layer (Qwen Sparse Attention), which selects relevant context elements, with three Gated DeltaNet layers, which compress history into fixed-size states. The second, called Gated Residual, uses four parallel residual branches with dynamic read/write gates, aiming to improve information flow and training stability. The third adds N-gram Embedding parameters, which allow consulting local token patterns at layer 2. According to the company, the 2.4T scale pre-training shows no loss spikes. In internal tests, Qwen4-Max-Base shows a bits per byte quantity 5 to 13% lower than Qwen3.8-Max-Base, while activating 25% fewer parameters.
Mixed Results Across Benchmarks
Alibaba’s comparative chart covers 16 benchmarks, pitted against Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, and Kimi K3.1. Several scores are taken from third-party rankings, and none had been independently reproduced at launch. Qwen4-Max ranks first on 4 of these 16 tests, notably HLAB-Vals-1.1 (legal agents), with 21.3 vs. 11.7 for Kimi K3.1 and 9.2 for Opus 5.5, as well as Agents’ Last Exam v1.1, with 46.7.
Claude Opus 5.5, however, dominates 7 of the 16 rankings. The most significant gaps are in code and vision: 31.5 on ProgramBench vs. 65.0 for Opus 5.5, and 48.5 on Terminal-Bench 4.0 vs. 67.2 for Kimi K3.1. On Work360, Alibaba’s internal benchmark covering 26 main domains and 721 specialized fields, the model scores 52.69, compared to 47.74 for Qwen3.8-Max.
Lower Usage Costs
According to Alibaba, token consumption decreases by approximately 24% at Xhigh effort and 35% at High effort on Work360. The average cost per task drops from $4.43 to $1.67, a 62% reduction. Public pricing follows the same trend: input costs $1.50 per million tokens, versus $2 for Qwen3.8-Max, and output costs $4.50, versus $6. Implicit cache reading drops by 70%, from $0.25 to $0.075.
The model carries the identifier qwen4-max on QwenCloud. Its API is compatible with OpenAI and Anthropic clients, and supports function calling, structured outputs, context caching, batch processing, web search, and fine-tuning. The default rate limit is 5,000,000 tokens per minute, and reasoning mode is activated with the enable_thinking parameter. According to Alibaba’s official post, the model weights are not released, ruling out self-hosting.
