
Alibaba Group has launched Qwen3-Next, a new AI model engineered for significant efficiency gains in training and inference, leveraging a highly sparse Mixture-of-Experts architecture. The base model, Qwen3-Next-80B-A3B-Base, achieves performance comparable to larger models while reducing training costs by over 90% and delivering more than tenfold throughput for long context lengths. This innovation, alongside specialized Instruct and Thinking versions excelling in ultra-long context and complex reasoning, positions Alibaba to offer powerful, cost-effective AI solutions, enhancing its competitive standing in the rapidly evolving AI landscape.
Alibaba has launched a new artificial intelligence model, Qwen3-Next, which represents a significant advancement in computational efficiency. The model's architecture, featuring a highly sparse Mixture-of-Experts (MoE) structure, allows the 80-billion parameter base model to activate only 3 billion parameters during inference. This results in performance comparable to the dense Qwen3-32B model but at less than 10% of the training cost in GPU hours, a material reduction in operational expenditure. For inference tasks with context lengths exceeding 32,000 tokens, the model delivers more than a tenfold increase in throughput, a critical performance metric for enterprise applications. The release of specialized versions, such as an 'Instruct' model that rivals Alibaba's flagship 235B-parameter model and a 'Thinking' model excelling in complex reasoning, demonstrates a strategic effort to capture diverse, high-value AI use cases. By making the model accessible through platforms like Hugging Face and the NVIDIA API Catalog, Alibaba is pursuing a broad adoption strategy that could drive utilization of its own cloud infrastructure and bolster its competitive standing in the global AI market.
AI-powered research, real-time alerts, and portfolio analytics for institutional investors.
Request a DemoOverall Sentiment
strongly positive
Sentiment Score
0.70
Ticker Sentiment