The Qwen team recently released its latest large language model, Qwen3, containing two MoE (Mixture of Experts) models and six dense models with parameter scales ranging from 0.6B to 235B. The flagship model, Qwen3-235B-A22B, performs well in a number of benchmarks including coding, math and generalization capabilities, and is competitive with top models such as DeepSeek-R1, o1, o3-mini, Grok-3 and Gemini-2.5-Pro.
The model is trained on 36T tokens covering 119 languages, with data sources including PDF extracts and synthetic data.Qwen3 also introduces thinking and non-thinking modes, improves agent and coding capabilities, and supports MCP (Model Control Protocol).Training pipeline similar to DeepSeek R1It also offers small distillation models such as the Qwen3-4B, which rivals even the Qwen2.5-72B-Instruct in performance, and the smaller Qwen3-0.6B model, which also performs well.
All models are open-sourced under the Apache 2.0 license, and users can experience them in the Qwen Chat Web and App, and visit platforms such as GitHub, Hugging Face, and ModelScope for more information.
- Qwen3-30B-A3B and its pre-trained version are now available on multiple platforms including Hugging Face, ModelScope and Kaggle.
- SGLang and vLLM frameworks are recommended for deployment.
- Suggested tools for local use include Ollama, LMStudio, MLX, llama.cpp, and KTransformers.
These options ensure that users can easily integrate Qwen3 into research, development or production environments.
---
Recently, Qwen 3 was released, a hybrid expert system (MoE) model with 235B parameters that demonstrates excellent performance with only 22B parameters when activated.
Qwen 3 excelled in several benchmarks and successfully outperformed well-known models such as R1, Grok and O1. This not only marks a major breakthrough for the Qwen series in the field of natural language processing, but also highlights its advantages of high performance and low resource consumption.
Technical Highlights:
- A hybrid expert system (MoE) architecture for efficient operation of large-scale models.
- The 22B activation parameter effectively reduces computational resource requirements.
In addition, Qwen 3 uses the open Apache 2.0 license, which allows developers to freely use and modify the code, greatly contributing to the development of the open source community and technological innovation.
This release is significant for both academia and industry, and is expected to drive the development of more applications based on Qwen 3, including, but not limited to, intelligent dialog, text generation, and machine translation.
---
The highly anticipated Qwen 3 model was recently released and its performance in a number of benchmarks was impressive.
As a new generation of large-scale language model, Qwen 3 has made significant progress in natural language processing, text generation and understanding. The model not only achieves breakthroughs in technology, but also shows strong potential in practical applications.
Qwen 3 The release of Qwen 3 has attracted widespread attention in the industry and its partner Ollama has been quick to follow suit by providing support for the new model. Users can now utilize Qwen 3 in Llama Index by following the steps below:
- Install the Llama Index Ollama plugin:
pip install llama-index-llms-ollama - Visit the official documentation: Detailed Guide
The launch of Qwen 3 marks a further development in language modeling technology, and both academic research and commercial applications will benefit from this powerful tool.
---
Qwen recently released six new Qwen 3 Series models, with the flagship modelQwen3-235B-A22BIt performs well in a number of benchmark tests.
This new batch of models not only provides better support in coding and multitasking (MCP), but also achieves significant competitive advantages against current top models such as DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro in a number of areas, including math and general-purpose capabilities. Particularly noteworthy is the fact that the small MoE model Qwen3-30B-A3B outperforms the QwQ-32B with only one-tenth of its activation parameters, while the even smaller Qwen3-4B even matches the Qwen2.5-72B-Instruct.
- Qwen3-235B-A22B: Flagship model with superior performance for complex missions.
- Qwen3-30B-A3B: Small MoE model with high parametric efficiency.
- Qwen3-4B: A lightweight model that is cost-effective.
To facilitate users' trial, these models are online on Qwen Chat Web and APP, and can be downloaded from GitHub, Hugging Face, ModelScope and other platforms. For deployment, the SGLang and vLLM frameworks are recommended, and for local use, tools such as Ollama, LMStudio, MLX, llama.cpp, and KTransformers are recommended. This makes it easy for users to integrate Qwen3 into their workflow, whether in research, development or production environments.
---
On April 29, 2025, Alibaba officially released the new generation of big language models, Tongyi Qwen3 ("Qwen3"). As the newest member of Alibaba's Tongyi series, Qwen3 has made breakthroughs in technical architecture, performance metrics, and deployment costs, especially in Agent capabilities and Model Context Protocol (MCP) support.
Thousand Questions 3 provides a total of 8 model versions, covering a wide range of needs from lightweight to ultra-large scale. Among them, the MoE (Mixed Expert) model includes two models, 30B (activation parameter 3B) and 235B (activation parameter 22B); and the Dense model offers a wide range of scales, including 0.6B, 1.7B, 4B, 8B, 14B, and 32B. The flagship model, Thousand Questions 3-235B-A22B, performs well in several international benchmarks, and is on par with international top-tier models such as DeepSeek-R1, OpenAI-o1, and Grok-3, and even slightly ahead on some tasks.
One of the notable features of Thousand Questions 3 is its significant improvement in Agent capability:
- In the BFCL benchmark test, which specializes in evaluating the Agent's task capabilities, Thousand Questions 3 reached a score of 70.8, surpassing high-end models such as Gemini 2.5 Pro and OpenAI-o1.
- Native support for tool calling (Function Calling) features, with the Qwen-Agent framework, providing pre-packaged tool calling templates and parsers, significantly reducing the development threshold, suitable for rapid development of AI Agent-based application scenarios.
In addition, Thousand Questions 3 supports a feature called Model Context Protocol (MCP):
- MCP enables the model to dynamically sense and manage contextual information, which significantly improves the stability and response accuracy of context perception when performing multi-round dialogues or complex tasks.
- Developers can clearly specify the "thinking budget" of the model, realize precise control, and flexibly meet the arithmetic cost and response quality requirements in different scenarios.
- The proposed hybrid reasoning model of "think fast" and "think slow", with dynamic context switching via MCP, allows the model to respond quickly to simple requests and automatically adopt an in-depth multi-step reasoning model for complex tasks.
ThousandQuest 3 is also quite attractive in terms of deployment: the flagship 235B model can be deployed with only 4 H20 GPUs, and the graphics memory footprint is only 1/3 of that of models of the same performance level, which significantly lowers the deployment cost threshold for enterprise users. Smaller models such as 4B and 8B can be easily deployed to end-side devices such as cell phones, PCs, and automobiles, making them suitable for various embedded and edge computing scenarios.
It is worth mentioning that Thousand Questions 3 supports 119 languages and dialects, including Chinese (Mandarin and Cantonese), English, Japanese, Korean, and a variety of Southeast Asian, European, and Middle Eastern languages, laying the foundation for internationalized applications and global market expansion.
Thousand Questions 3 series models are open source along the Apache 2.0 protocol, and developers worldwide are free to download, modify, and commercialize them. The platforms provided include Hugging Face, ModelScope, Kaggle and so on. Alibaba's Magic Match community and Alibaba Cloud's Hundred Refinement Platform also provide APIs and deployment support.
---
In just over two years, Llama has surpassed one billion downloads, establishing itself as a leader in the AI open source ecosystem.
During the LlamaCon 2025 keynote, Llama announced a number of important updates and developments:
- Llama API PreviewLlama: Combines the best features of the Closed Model API with the flexibility of open source to provide developers with more powerful tools. the Llama API supports a wide range of programming languages and provides rich documentation and sample code to help developers get started quickly.
- AI Defender Program: Introduces advanced tools for detecting and preventing AI-generated threats. The program will leverage the latest machine learning technologies to help businesses protect their data security and privacy.
- Llama Impact Grant Winners Announced: Introducing winners who are driving change through open source AI technology. These projects cover a wide range of fields, from healthcare to environmental protection, and demonstrate the enormous potential of open source AI to solve real-world problems.
---
Chinese tech giant Alibaba recently released Qwen 3, the latest generation of its big language model, which demonstrates excellence in a number of ways.
Qwen 3 excels in several tests, especially in reasoning and memorization abilities that far exceed those of the currently popular Llama model. Specifically.Qwen 3 has 8 fully open-source model versionsand are all licensed under the Apache 2.0 license, which provides great flexibility and convenience for developers and researchers.
- Stronger reasoning: Qwen 3 excels at complex tasks and can easily solve a wide range of puzzles.
- Improved Memory Mechanisms: Compared to Llama, Qwen 3 has better contextual understanding and long text memorization.
- Built-in hybrid inference: This unique design makes Qwen 3 more efficient when dealing with multimodal data.
In addition, Qwen 3 demonstrates its powerful performance in real-world applications. For example, Qwen 3 has demonstrated significant advantages in areas such as natural language processing, code generation, and dialog systems. These advances not only improve the overall performance of the model, but also lay a solid foundation for future AI technology development.
Overall, the release of Qwen 3 marks another major breakthrough in the field of large language modeling in China, and its performance advantages are already very obvious.
---
Recently, a program called8BThe multilingual processing model has attracted a lot of attention. The model not only supports multiple languages, but also has code processing and proxy capabilities, and is released under the Apache license, which means developers are free to use and modify it.
It is worth noting that8BThe model is compatible with the popular Transformer architecture as well as the Llama.cpp library, providing developers with great flexibility and extensibility. In addition, due to its modest size, the model can run on most laptops and MacBooks without the need for high performance hardware support.
In practice.8BThe model can be used for natural language processing tasks such as text generation, translation and dialog systems. It is also capable of handling code-related tasks and providing support for programming aids. For projects that require multilingual support and code processing capabilities, the8BIt is a very attractive option.
---
Alibaba has officially released the latest generation of its Qwen3 family of models, which consists of 8 open-weighted models covering 6 intensive models (0.6B, 1.7B, 4B, 8B, 14B, 32B) and 2 hybrid expert system models (30B-A3B, 235B-A22B). These models perform well on coding, mathematical, and generalized tasks, and are able to compete with top models such as DeepSeek-R1 and o1.
The Qwen3 family of models is pre-trained with about 36 trillion tokens and supports 119 languages and dialects with enhanced autonomy. Users can switch between thinking and non-thinking modes to adapt to different application scenarios.
Key Features:
- Support for 119 languages and dialects
- Pre-training data size of 36 trillion tokens
- Provides two modes: thinking mode and non-thinking mode
- Based on the Apache 2.0 license open source
The Qwen3 family of models now fully supports the BentoML and OpenLLM frameworks, allowing users to easily deploy these models on BentoCloud. In addition, BentoML and vLLM provide a convenient way to serve models, further simplifying their use and management.
By following the link below, you can quickly experience the power of the Qwen3 model:
235B-A22B: https://t.co/84C77s9lbg
8B: https://t.co/u42LpNqNuX
For more information, please visit: https://t.co/YeawhsoREX
For more information about OpenLLM, please visit: https://t.co/PVhLM89LSc
---
At the recent LlamaCon conference, Meta announced a series of major AI safety and protection tool updates designed to enhance the safety of AI systems through open source tools and technologies.
These updates include:
- Llama Guard 4: A customizable security protection tool that supports cross-modal text and image comprehension protection.
- Llama Firewall: A security fence tool that detects and prevents threats such as hint injection, unsafe code and risky LLM plugin interactions.
- Two new versions of theLlama Prompt Guard::
- Prompt Guard 2 86M: Improved performance for jailbreak and prompt injection detection.
- Prompt Guard 2 22M: Smaller and faster, reducing latency and computational costs while maintaining high performance.
In addition, Meta has invested in the development of new AI-enabled solutions to help communities enhance their security systems. Among them.CyberSecEval 4is the latest set of cybersecurity benchmarking tools for AI systems.
Also.Llama Defender ProgramA wide range of open, early access and closed AI solutions will be available for trusted partners to address different security needs.
For more information on these open source protection tools and how they advance AI privacy and security, visit: the official LlamaCon website
