Introducing Qwen3: The Next Frontier in AI Language Models

In the ever-evolving landscape of artificial intelligence, Alibaba Cloud's Qwen team has once again pushed the boundaries with the release of Qwen3, their latest and most advanced large language model to date. This groundbreaking addition to the Qwen family promises to redefine the capabilities of AI systems, offering remarkable improvements across a wide range of tasks, from logical reasoning and code generation to commonsense reasoning and multilingual instruction following.

We are excited to announce the release of Qwen3, the latest addition to the Qwen family of large language models.

Scaling New Heights with Dense and Mixture-of-Expert Models

One of the standout features of Qwen3 is its availability in both dense and Mixture-of-Expert (MoE) model architectures, catering to a wide range of AI applications and computational requirements. The dense models range from 0.6B to 32B parameters, while the MoE models span from 30B-A3B to an impressive 235B-A22B parameters, making Qwen3 one of the largest and most capable language models to date.

The highlights from Qwen3 include: Dense and Mixture-of-Experts (MoE) models of various sizes, available in 0.6B, 1.7B, 4B, 8B, 14B, 32B and 30B-A3B, 235B-A22B.

Enhanced Reasoning and Code Generation Capabilities

One of the key areas where Qwen3 shines is in its improved logical reasoning and code generation abilities. By leveraging advanced language understanding and generation techniques, Qwen3 can tackle complex reasoning tasks with greater accuracy and efficiency, making it an invaluable tool for developers, researchers, and anyone working with code or logical systems.

I have 128GB VRAM + 192GB RAM (consumer motherboard, 7800X3D at 6000Mhz, so just dual channel), and depending of offloading some models can have pretty decent speeds. Qwen 235B at Q6\_K, using all VRAM and \~70GB RAM I get about 100 t/s PP and 15 t/s while generating. DeepSeek V3 0324 at Q2\_K\_XL using all VRAM and \~130GB RAM, I get about 30-40 t/s PP and 8 t/s while generating.

Multilingual Prowess and Human-Aligned Conversations

Qwen3 also excels in its multilingual capabilities, supporting over 100 languages and dialects, making it a valuable asset for businesses and organizations operating in diverse linguistic environments. Additionally, the model boasts enhanced human preference alignment, enabling more natural and engaging conversational experiences that feel tailored to individual users.

i definitely hear him talk like tony soprano

Seamless Switching Between Thinking and Non-Thinking Modes

One of the innovative features introduced in Qwen3 is the ability to seamlessly switch between thinking and non-thinking modes, optimizing performance across different scenarios. This functionality allows the model to adapt its behavior based on the task at hand, ensuring efficient resource utilization and optimal results.

Comprehensive Documentation and Integration Support

To facilitate the adoption and deployment of Qwen3, Alibaba Cloud has provided comprehensive documentation covering various aspects of the model, including quickstart guides, inference instructions, local running, deployment, quantization, and training instructions. Additionally, the documentation details integration with various frameworks and tools, ensuring seamless integration into existing workflows and pipelines.

Pushing the Boundaries of Open-Source AI

With the release of Qwen3, Alibaba Cloud's Qwen team aims to provide the public with access to advanced AI models that excel in a wide range of tasks, from creative writing to complex agent-based tasks. By embracing open-source principles, the team hopes to foster collaboration and accelerate the development of AI technologies, ultimately pushing the boundaries of what open-source models can achieve.

https://preview.redd.it/71w2erb9ryye1.jpeg?width=1182&format=pjpg&auto=webp&s=09a450824d1ab28aa506a4425e990186ae7b53fa My chat is never the fun one :(

Conclusion

The release of Qwen3 marks a significant milestone in the development of large language models, showcasing Alibaba Cloud's commitment to pushing the boundaries of AI technology. With its impressive capabilities across reasoning, code generation, multilingual support, and human-aligned conversational experiences, Qwen3 is poised to revolutionize the way we interact with and leverage AI systems. As the open-source community embraces this powerful tool, we can expect to witness a surge of innovative applications and breakthroughs that will shape the future of artificial intelligence.