
Tech • AI • Robotics • Game
Alibaba has launched Qwen 4 Max, a flagship multimodal model with strong web development and agentic workflow features, but early independent testing suggests it trails several top rivals on overall performance and value.
Qwen 4 Max is the first released model in the Qwen 4 family and is positioned as a general-purpose multimodal system for text, image and video understanding. It supports advanced reasoning, web search, tool calling, structured outputs and long-horizon agent workflows. Alibaba says it is built on a new mixture-of-experts architecture with native vision and a 1 million-token context window.
Alibaba says the model improves on Qwen 3.8 Max while using up to 50% fewer tokens and cutting average task cost by more than 60% in internal evaluations. On third-party tracking cited in the release discussion, it scored 43 on the Artificial Analysis Intelligence Index, a 9-point gain over Qwen 3.8 Max. Reported output token usage was also down 39% from the prior flagship.
The model is available through Qwen Chat and via API on Alibaba’s cloud platform. Pricing was listed at $1.50 per million input tokens and $4.50 per million output tokens. Per-task estimates cited in the discussion put usage at about $3.21 on one benchmark methodology, making it cheaper than its predecessor but still not clearly ahead of all close competitors on value.
Despite the upgrade, Qwen 4 Max appears to sit behind several leading models in outside comparisons. In one benchmark cited, it ranked below GPT-6 Soul, GPT-6.1 Terra, Opus 5 and Kimi K 3.1 Pro. The same discussion noted that on the Artificial Analysis index it is 12 points behind GPT-6.1 Soul while costing more than twice as much per task, raising questions about its intelligence-per-dollar position.
The model shows clear strength in front-end generation but less convincing results in broader coding tests. On Code Arena, it reportedly reached the No. 6 spot in web development, suggesting strong performance on UI-heavy tasks. But on Terminal Bench 4.0 it scored 48%, well behind Kimi’s 67%, indicating weaker performance in terminal-style coding and execution tasks.
Web creation is one of the model’s clearest advantages. It was credited with following detailed design instructions more reliably than earlier Qwen models, producing front ends with custom typography, scroll behavior and consistent visual style. A MacOS-style clone included functioning interface elements such as a top bar, app windows, animations, calculator, stocks panel and a music player.
The model was also used to generate playable game projects with engines such as Unity, Unreal Engine and Godot. One Unity example included visual effects, audio, character animation, HUD elements and gameplay systems. Another task produced a Minecraft Dungeons-style action game with multiple floors, enemy encounters and environment changes, showing that the model can assemble large interactive projects with few visible errors.
Speed is a recurring concern. A single game-generation task reportedly took about 2 hours, while a technical simulation of GPU, tensor core and TPU behavior required six refinement rounds and roughly a day to complete. That leaves the model at a disadvantage against faster systems that can deliver similar or better outputs with lower token use.
Alibaba highlighted the model’s performance on professional tasks in finance, law and healthcare. In one finance test, it reportedly used 90 MCP tools to conduct investment research and generate traceable Excel and PowerPoint outputs. In legal work, it organized 198 documents into an 86-page due diligence report that flagged 14 legal issues. In healthcare, it processed about 1,400 medical images, connected them with patient records and generated recommendations said to align with clinical decisions.
While Qwen 4 Max itself is not open source, future Qwen 4 releases are expected to include open-weight models. The anticipated 27 billion-parameter version could prove especially significant if it retains much of the flagship’s capability while being small enough for local deployment. That prospect may be more compelling to developers than the current top-end release.
Qwen 4 Max expands Alibaba’s multimodal and agentic AI stack with notable strengths in web development and professional workflows. But slower responses, mixed coding results and weaker value against top competitors make the launch look more incremental than transformative.
Ask a question