Alibaba open-sources Qwen3-Next with 10x the power at one-tenth the cost

Alibaba Group Holding has unveiled its latest artificial intelligence model, marking a major leap in efficiency and cost-effectiveness. The model, developed under the Qwen3-Next architecture, is said to be 10 times more powerful yet requires only a tenth of the cost to build compared with its predecessor, according to announcements made Friday by the company’s Qwen team.

A Cost-Efficient Powerhouse

The new flagship, Qwen3-Next-80B-A3B, boasts 80 billion parameters. In benchmarking tests, it not only outperformed the earlier Qwen3-32B, released in April, but also matched the performance of the company’s top-tier Qwen3-235B-A22B model. Despite its lower cost, the 80B model demonstrated up to 10-fold faster performance in certain tasks.

Alibaba Cloud, the AI and cloud computing arm of Alibaba, emphasized that these models are optimized for efficient deployment on consumer-grade hardware, signaling a shift toward broader accessibility in advanced AI development.

Open-Source Strategy to Narrow the Gap

The release underscores a growing trend among Chinese AI firms to leverage open-source frameworks as a way to close the gap with U.S. rivals. By making its models publicly available on GitHub and Hugging Face, Alibaba Cloud is fostering one of the largest open-source AI ecosystems in the world.

This open approach allows third-party developers to freely use, adapt, and distribute Alibaba’s models, further accelerating innovation across industries.

Advancing Reasoning Capabilities

Alongside the base model, Alibaba launched a reasoning-focused version under the same architecture, known as Qwen3-Next-80B-A3B-Thinking. Benchmark results showed it outperformed both Alibaba’s own Qwen3-32B-Thinking and Google’s Gemini-2.5-Flash-Thinking across multiple reasoning tasks.

Apple Integration and MLX Compatibility

The Qwen3 family is also notable for its integration with Apple’s MLX framework, enabling developers to train and run AI models on devices like the iPhone. This follows earlier reports that Apple partnered with Alibaba to use Qwen models for Apple Intelligence in mainland China, while relying on OpenAI’s GPT models internationally.

Architectural Innovations Driving Efficiency

The Qwen3-Next efficiency gains stem from several technical breakthroughs:

• Hybrid Attention: Improves handling of long text inputs.
• High-Sparsity Mixture-of-Experts (MoE): Distributes tasks across specialized sub-networks for better balance between performance and efficiency.
• Multi-Token Prediction: Speeds up inference and boosts training outcomes.
• Enhanced Training Stability: Improves consistency across training runs.

The MoE framework, in particular, allows the model to scale intelligently without the steep computational burden that typically accompanies larger AI systems.

Scaling to the Trillion-Parameter Frontier

The launch comes just a week after Qwen-3-Max-Preview, Alibaba’s largest AI model yet, featuring over 1 trillion parameters. While massive models typically deliver stronger capabilities, they also demand immense computing power. Nevertheless, Qwen-3-Max-Preview debuted at sixth place on “text arena,” a ranking platform run by researchers at the University of California, Berkeley.

Outlook

Alibaba Cloud’s rapid innovation in the Qwen3 series highlights both its ambition and its growing technological competitiveness. By combining cutting-edge architecture with open-source accessibility, the company is strengthening its position in the global AI race while empowering developers worldwide.