Alibaba has released its most powerful AI model yet, the Qwen3-Max, a system with more than 1 trillion parameters designed to propel China’s tech giant deeper into the global AI race.
The record-breaking model was introduced at the company’s annual conference. Alibaba Cloud Chief Technology Officer Zhou Jingren told Reuters it promises advanced code generation and agentic capabilities.
Trained on massive data sets
The Qwen3-Max was trained on 36 trillion tokens and can process inputs as long as one million words or symbols — roughly the length of several books — far beyond most AI systems today.
Its Mixture of Experts design was built to maintain steady performance, avoiding the crashes and resets that often occur with training ultra-large models. Alibaba noted the training ran smoothly without loss spikes or rollbacks.
The company states that it has also refined the training process with new methods, making it more efficient and reliable. One approach, called ChunkFlow, tripled the system’s speed when working with long inputs, while overall training efficiency improved by 30% compared to the previous Qwen2.5.
Other safeguards, including tools named SanityCheck and EasyCheckpoint, cut downtime from hardware failures to a fraction of what earlier models faced.
Moving toward agentic AI
Qwen3-Max’s agentic abilities enable it to take on tasks with fewer human prompts and move toward goals independently, a step beyond traditional chatbots.
Two versions are being rolled out. The Instruct model is already live on Alibaba Cloud and Qwen Chat, while a more advanced “Thinking” version is still in training. Developers can plug into Qwen3-Max through APIs that are compatible with OpenAI’s, making it easier for those who have worked with other leading models to build on Alibaba’s platform.
Topping global leaderboards in coding and reasoning
Qwen3-Max is already placing near the top of international AI rankings. The Instruct version secured a top-three global spot on the Text Arena leaderboard, edging out OpenAI’s GPT-5-Chat.
On SWE-Bench Verified, a test built around solving real-world programming problems, it scored 69.6, putting it among the strongest coding models to date.
The system also excelled in agent benchmarks. On Tau2-Bench, which measures how well AI models can use tools and carry out multi-step tasks, Qwen3-Max reached 74.8, outperforming competitors such as Claude Opus 4 and DeepSeek V3.1. Alibaba has highlighted these results as proof of the model’s versatility beyond simple conversation.
Alibaba doubles down on AI investment
Earlier this year, Alibaba pledged 380 billion yuan ($53.4 billion) over three years to develop AI infrastructure, ranging from data centers to training platforms. At the event, Chief Executive Eddie Wu announced that the company would increase spending further, although he did not provide a specific figure, reflecting Alibaba’s push to make artificial intelligence a core business pillar.
Wu added that demand for advanced systems is already surpassing expectations.
“The speed of AI industry development has far exceeded our expectations, and the industry’s demand for AI infrastructure has also far exceeded our expectations,” he told Reuters.
Expanding into multimodal systems
Alibaba also unveiled Qwen3-Omni, a multimodal system for immersive applications such as smart glasses and intelligent cockpits. The move indicates that Alibaba is targeting not only researchers and developers but also aiming to integrate AI into consumer devices and everyday use.
The dual focus on infrastructure and consumer products shows Alibaba’s determination to broaden its role in the global AI push.
Chipmakers are racing to design processors that replicate the brain’s efficiency in handling information. See how this emerging approach could set the stage for the next generation of AI hardware.
The post Alibaba’s ‘Most Powerful’ AI Model: Qwen3-Max Packs 1 Trillion Parameters appeared first on eWEEK.
No Responses