Qwen3-Next - the latest base model from Ali Tongyi

Latest AI Resources3dys agorelease AI Sharing Circle

4.3K 00

What is Qwen3-Next

Qwen3-Next is a new generation of hybrid architecture of Ali Tongyi open source large model, combining Gated DeltaNet and Gated Attention technology, good at processing long text, reasoning fast and save computing resources. The model is divided into a command version (Qwen3-Next-80B-A3B-Instruct ) and a thinking version (Qwen3-Next-80B-A3B-Thinking), which are good at understanding and executing commands and deep reasoning respectively. The total number of parameters in the model reaches 80B, and each inference only activates about 3B parameters, which significantly reduces the computational cost.Qwen3-Next has a wide range of applications in many fields, such as intelligent customer service, content generation, data analysis, etc., and can provide users with efficient and accurate services. VisitAlibaba's hundreds of refinementscap (a poem)QwenChatThe web version can be experienced.

Features of Qwen3-Next

Instruction Understanding and Execution: Understands user commands accurately and executes them efficiently for a wide range of automation tasks.
in-depth reasoning ability: Supports complex multi-step reasoning and is suitable for solving problems that require deep thinking.
Long Text Processing: Supports processing of very long text (32K or more), suitable for processing large amounts of information.
Efficient Reasoning: Based on a hybrid engine with fast inference and low resource consumption.
Saving resources: Only about 3B parameters are activated per inference, significantly reducing computational cost.
Multi-disciplinary applications: Apply to a variety of fields such as intelligent customer service, content generation, data analytics, educational assistance, and legal counseling.

Core Benefits of Qwen3-Next

hybrid architecture: Combining Gated DeltaNet and Gated Attention to achieve a balance of speed and accuracy.
Long Text Processing: Supports efficient processing of 32K+ long texts, suitable for scenarios where a large amount of text information needs to be processed.
efficient reasoning: Inference is more than 10 times faster than Qwen3-32B for applications that require fast response time.
Resource savings: With 80B total number of parameters, only about 3B parameters are activated per inference, significantly reducing computational cost.
Expert system (MoE): Contains 512 experts and dynamically selects the most relevant experts to work with for load balancing.
Pre-training acceleration: Reduce the number of inference steps and increase the speed of long text generation through native MTP acceleration technology.

Performance of Qwen3-Next

Instruct Model Performance: Qwen3-Next's Instruct model has demonstrated superior instruction comprehension in several benchmarks, on par with the 235B flagship model. In terms of long text processing, it takes advantage of its unique architecture to process and analyze large amounts of text more efficiently, ensuring the completeness and accuracy of information.
Thinking Model PerformanceQwen3-Next's Thinking model excels in reasoning ability, surpassing Gemini Flash. in some key indicators, it is even close to the level of the flagship model of 235B, demonstrating powerful multi-step reasoning and deep thinking ability, capable of dealing with complex logical problems and providing accurate solutions.

What is Qwen3-Next's official website?

HuggingFace Model Library:: https://huggingface.co/collections/Qwen/qwen3-next-68c25fd6838e585db8eeea9d

People for whom Qwen3-Next is suitable

Customer Service Team: The model responds quickly to customer inquiries, provides 24/7 automated services, and improves customer satisfaction.
Content Creation Department: In copywriting, article creation, advertising planning, etc., the ability to quickly generate high-quality text content and improve creative efficiency.
Data Analytics Team: Models help analyze large amounts of textual data and extract key information to support decision-making.
Product Development Team: Provide assistance in product requirement analysis, user experience optimization, etc. to help the team better understand user needs.
principals: Tutorials can generate content, design course syllabi, write lesson plans, etc., reducing the burden of lesson planning.