What is Qwen3-Next
Qwen3-Next is a new-generation hybrid architecture large model open-sourced by Alibaba Tongyi, including two versions: the Instruct version (Qwen3-Next-80B-A3B-Instruct) and the Thinking version (Qwen3-Next-80B-A3B-Thinking). The model uses the world's first Gated DeltaNet and Gated Attention hybrid architecture to achieve faster processing speed, stronger performance, and lower resource consumption. The model has a total parameter count of 80B, with only about 3B parameters activated per inference. The model's pretraining acceleration mechanism significantly improves long-text generation speed. Qwen3-Next is suitable for long-context reasoning, high-concurrency, and low-latency scenarios, and the model is now available onAlibaba Cloud Bailiansupports API calls andQwenChatthe web version supports online experience.
Qwen3-Next's Main Features
- Instruction understanding and execution: The specially optimized Instruct version is good at understanding and executing user instructions.
- Deep thinking and reasoning: The Thinking version supports multi-step reasoning and deep thinking about problems.
- Long text processing: The model can process and understand long text inputs of more than 32K.
- High-efficiency inference: Based on a hybrid engine, it combines speed and precision to achieve fast and accurate inference.
- Resource saving: While maintaining large model capabilities, it saves computational resources by activating only part of the parameters.
Qwen3-Next's Technical Principles
-
Hybrid architecture: Combines 75% Gated DeltaNet and 25% Gated Attention to balance speed and precision.
-
Gated DeltaNet: Used for processing long text, providing fast inference speed and linearly growing memory usage.
-
Gated Attention: Responsible for accurately recalling key information to ensure the model does not lose focus when processing long text.
-
-
Extreme cost-saving architecture: The model has a total parameter count of 80B, and only about 3B parameters are activated per inference, significantly reducing computational cost.
-
Pretraining acceleration mechanism: Through native MTP (Multi-Token Prediction) acceleration technology, the model can predict multiple tokens at once during the pretraining stage, reducing inference steps when generating long text and improving throughput.
-
Expert system (MoE): The model contains 512 experts, and when processing each request it dynamically selects the 10 most relevant experts and 1 shared expert to work, achieving load balancing and resource optimization.
Qwen3-Next's Project Address
- HuggingFace Model Repository：https://huggingface.co/collections/Qwen/qwen3-next-68c25fd6838e585db8eeea9d
Qwen3-Next's Performance
- Instruct model performance: In multiple benchmark tests, the instruction capability of the Qwen3-Next-80B-A3B-Instruct model is comparable to235Bflagship models, and it performs even better in processing long text.
- Thinking model performance: The Qwen3-Next-80B-A3B-Thinking model surpasses Gemini Flash in reasoning ability, with some metrics even approaching the 235B flagship model.
How to Use Qwen3-Next
- Alibaba Cloud Bailian: VisitAlibaba Cloud Bailianofficial website, search for Qwen3-Next, and call the model according to the provided API reference.
- QwenChat web version: Directly visitQwenChatofficial website to interact with the model online.
- Hugging Face: On Hugging Face, find the Qwen3-Next model collection to call and experience the model.
Application Scenarios of Qwen3-Next
-
Intelligent Customer Service: Utilizing the model's instruction understanding and execution capabilities to provide automated customer service and support.
-
Content Generation: In fields such as news, blogs, and social media, it can automatically generate articles, stories, or other text content.
-
Data Analysis: Data analysts analyze large amounts of text data, extract key information, and support market research and business intelligence.
-
Educational Assistance: In the education field, it assists students in learning and provides personalized teaching content and tutoring.
-
Legal Consultation: Legal professionals analyze legal documents and provide legal consultation and research support.