I've been using it today and it's without a doubt the most useful model in the opensource scene at yhe moment. The 30b Moe is really fast for consumer grade GPUs
Brief:
- Qwen 3 is an open-source model that rivals Gemini 2.5 Pro in performance.
- The flagship Qwen 3 model (235B) with 22 billion active parameters outperforms Gemini 2.5 Pro in several benchmarks, especially in coding and function calling.
- Qwen 3 introduces a hybrid thinking model, offering both deep reasoning (thinking mode) and fast responses (non-thinking mode), allowing users to adjust the thinking budget.
- The model is optimized for agent-based tasks like coding and is compatible with MCP tools, making it a strong choice for automation and integration.
- Qwen 3 features models ranging from 600 million to 235 billion parameters, with a focus on efficiency and tool integration.
- It supports tool calling during problem-solving, enabling complex tasks like code execution and file management within the same inference run.
- Qwen 3 was trained on 36 trillion tokens, including a mix of web data, PDFs, textbooks, and synthetic data to enhance math and coding knowledge.
- The training pipeline for Qwen 3 included stages for reasoning, reinforcement learning, and non-thinking capabilities, making it highly versatile.
- Qwen 3 outperforms competitors like Llama 4 in various benchmarks, with fast execution times and strong general task performance.
- The model is available for download and can be tested through platforms like LM Studio, with promising results in coding tasks like writing Python games.