
Alibaba's Qwen4-Max challenges Opus 5.5, GPT-6 and Kimi K3.1 Pro at lower prices
Alibaba launched Qwen4-Max, a flagship model with 2.4 trillion parameters that activates 71 billion per token and offers a one-million-token context window at much lower task costs.
Alibaba has launched Qwen4-Max, a new flagship model that pairs a more efficient architecture with lower prices and targets professional work in industry, science and finance. The mixture-of-experts model has 2.4 trillion parameters but activates only 71 billion per token, processes images natively and offers a one-million-token context window. It is positioned against Opus 5.5, GPT-6 and Kimi K3.1 Pro.
Architecture and training
Alibaba combines standard attention with a hybrid method: three layers compress prior context into a condensed state, and a fourth layer selects only important parts, cutting compute on long inputs. The design adds four parallel connections with dynamic gates and an N-gram layer with 328 billion parameters. Alibaba says training ran stably, with loss falling steadily and no spikes. Internal bits-per-byte tests found Qwen4-Max-Base best among five checkpoints, improving 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.
Cost and benchmarks
On Work360 across 721 fields, Qwen4-Max uses roughly a quarter to a third fewer tokens than its predecessor, cutting average cost per task by 62 percent from $4.43 to $1.67 while its score rises from 47.74 to 52.69. Terminal-Bench 4.0 token use drops by nearly half. On the Artificial Analysis Intelligence Index, it scores 43 points, nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro at 46; cost is $3.27 per task, about a third of Qwen3.8-Max's $10.02 and roughly 11 percent less than Kimi K3.1 Pro's $3.68, helped by 25 percent lower token prices, 70 percent cheaper cache hits and fewer tokens. External results are mixed: it leads on Agents' Last Exam v1.1 at 46.7 points, trails Opus 5.5 and GPT models on Humanity's Last Exam, and falls behind GPT models but edges past Opus 5.5 on HealthBench Professional. It scores 48 percent on Terminal-Bench 4.0 versus Kimi K3.1 Pro's 67 percent, ties Kimi at 26 percent on FrontierCode 2.0 and trails Tencent Hy4 at 31 percent. API pricing is $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for cache hits.
Agents, labs and games
Live Delivery provides a usable first version early and builds it out later, letting users step in and change course at any point; Alibaba open-sourced the Qwen-Live-Delivery-Plugin. SANF and Pacific Union cut manual programming time for a single workpiece from about 475 minutes to 24, while finance use taps 90 MCP tools and returns verifiable Excel and PowerPoint files with traceable numbers. Internally, Qwen4-Max with MindOpt cut data-annotation compute allocation costs by an average of 10.42 percent and found optimal CPU task distributions more than 14 times faster. In lab tests, it ran autonomously for 48 hours with Mira from Deep Principle, searching 2.7 million molecular candidates, and optimized 24 candidate molecules with PharmMaster from DP Technology, with the final candidate binding its target protein about 3.6 times better than a known starting compound. Alibaba also shows single-prompt game development across Unity, Unreal Engine and Godot, including 'Meow of Magic,' and reusable 3D models in Blender.
Safety and availability
On Last Secure Code Bench, code that is both functional and secure rose from 22 to 31.8 percent, and the share secure at all three testing levels nearly doubled from 8 to 14.7 percent. Alibaba says reviewers found no real capacity for innovation that would give users a significant capability boost on CBRN risks. Qwen4-Max is available through QwenCloud and in QwenWork, Qoder and QwenStudio, with smaller Qwen4 models planned as open source later.
Sources: The Decoder · The New Stack
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.