Skip to main content
Alibaba Qwen4-Max AI model details: 2.4T MoE, 71B active parameters, 1M context, 5-13% lower bits/byte.

Editorial illustration for Alibaba Qwen4-Max: 2.4T MoE, 71B active, 1M context, 5-13% lower bits/byte

Alibaba Qwen4-Max: 2.4T MoE Model, 1M Context

• 4 min read

Alibaba's Qwen team put out Qwen4-Max this month, the first release in the Qwen4 line and a sharp break from the open-weight habit that made Qwen a go-to name for self-hosters. This one stays locked behind Alibaba's own infrastructure: QwenCloud API, Token Plan, Qwen Studio, QwenWork, Qoder. No weights, no quantized builds, no running it on your own GPUs.

The specs explain some of that caution. Qwen4-Max is a sparse mixture-of-experts model with a 2.4 trillion parameter backbone, but only 71 billion of those parameters activate per token, alongside a separate 328 billion parameter N-gram embedding component. It reads text, images and video, and holds a 1 million token context window with up to 393,000 tokens of output. Alibaba paired the launch with a price cut, claiming a 62% drop in average cost per task compared to Qwen3.8-Max.

The architecture changes target three spots in the transformer stack: attention, residual connections and embeddings. Alibaba frames the result as a jump over its predecessor, competitive with frontier models on agentic tasks, though it lags on hard coding and vision benchmarks.

Two things stack: fewer tokens and lower prices. On Work360, token use drops about 24% at Xhigh effort and 35% at High effort. On Terminal-Bench 4.0, the drops are about 48% and 49%.

Why this matters A 2.4-trillion-parameter backbone that only activates 71 billion tokens at a time, plus 328 billion N-gram embedding parameters, is Alibaba's bet that sparsity can keep scaling without blowing up inference costs. The bits-per-byte gain over Qwen3.8-Max-Base, 5 to 13% better while using 25% fewer active parameters, is a real efficiency story, and the 62% drop in average task cost is the number developers will actually feel in their API bills. For teams building on long documents or video, the 1M-token context and 393K output ceiling open workflows that were previously awkward to chunk.

But first place on only 4 of 16 benchmarks against Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 is a mixed scorecard, not a sweep. Alibaba is clearly optimizing for cost-per-task rather than chasing every leaderboard, which is a reasonable strategy if you're selling API volume. Worth watching: independent verification of those in-house numbers, and how the remaining 12 benchmarks break down before anyone calls this a clear leader.

Common Questions Answered

What is the key difference between Qwen4-Max and previous Qwen models in terms of deployment?

Unlike earlier Qwen models that were open-weight and could be self-hosted, Qwen4-Max is proprietary and only accessible through Alibaba's infrastructure including QwenCloud API, Token Plan, Qwen Studio, QwenWork, and Qoder. This represents a significant departure from Alibaba's previous open-weight approach that made Qwen popular among self-hosters.

How does Qwen4-Max's mixture-of-experts architecture reduce inference costs?

Qwen4-Max uses a sparse mixture-of-experts design with a 2.4 trillion parameter backbone but only activates 71 billion parameters at inference time, significantly reducing computational overhead. This sparsity approach achieves 5-13% better efficiency in bits-per-byte compared to Qwen3.8-Max-Base while using 25% fewer active parameters, resulting in a 62% drop in average task cost for developers.

What is the context window size for Qwen4-Max and why does it matter?

Qwen4-Max supports a 1 million token context window, enabling it to process significantly longer documents and conversations than models with smaller context limits. This extended context capability is particularly valuable for teams building applications that require understanding and analyzing large volumes of text data.

How much do token usage and costs decrease when using Qwen4-Max compared to previous versions?

Token usage drops approximately 24% at Xhigh effort and 35% at High effort on Work360 benchmarks, with even more dramatic reductions of 48-49% on Terminal-Bench 4.0. These token reductions directly translate to lower API costs, making Qwen4-Max more economical for developers using Alibaba's infrastructure.

What role do N-gram embedding parameters play in Qwen4-Max's architecture?

Qwen4-Max incorporates 328 billion N-gram embedding parameters as part of its overall architecture, working alongside the sparse mixture-of-experts design to optimize model performance and efficiency. These parameters contribute to the model's ability to achieve superior bits-per-byte efficiency while maintaining strong performance across various tasks.

LIVE13:40Alibaba Qwen4-Max: 2.4T MoE, 71B active, 1M context, 5-13% lower bits/byte