Qwen3 was developed by the Alibaba Cloud Qwen team and is part of the Qwen large language model series. The materials describe Qwen3 as one of the most advanced and intelligent systems currently available, building on the experience of the earlier QwQ and Qwen2.5, and making the model weights publicly available.
Main capabilities and model scales: Qwen3 offers both Dense and Mixture-of-Experts (MoE) models, with scales including 0.6B, 1.7B, 4B, 8B, 14B, 32B, as well as MoE models 30B-A3B and 235B-A22B. Its features include the ability to seamlessly switch between thinking mode (for complex logical reasoning, mathematics, and programming) and non-thinking mode (for efficient general-purpose dialogue); the official description states improvements over previous models in mathematics, code generation, and commonsense logical reasoning; relatively good performance in human preference alignment such as creative writing, role-playing, multi-turn dialogue, and instruction following; agent capabilities, enabling integration with external tools in both thinking and non-thinking modes; support for more than 100 languages and dialects, with multilingual instruction following and translation capabilities.
Qwen3-2507 update: The materials mention that the potential of the Qwen3 series was explored over the past three months, and two variants of Qwen3-2507 were introduced: Qwen3-Instruct-2507 and Qwen3-Thinking-2507, along with three scales: 235B-A22B, 30B-A3B, and 4B. Qwen3-Instruct-2507 is an updated version of the previous Qwen3 non-thinking mode. Its main enhancements include: significant improvements in general capabilities such as instruction following, logical reasoning, text understanding, mathematics, science, programming, and tool use; improved coverage of multilingual long-tail knowledge; better alignment with user preferences in subjective and open-ended tasks, generating more helpful responses and higher-quality text; enhanced 256K token long-context understanding, expandable to 1 million tokens. Qwen3-Thinking-2507 is a continuation of the Qwen3 thinking model, improving reasoning quality and depth: improved performance on logical reasoning, mathematics, science, programming, and academic benchmarks that usually require human expertise, with the official description stating that it achieves state-of-the-art among open-weight thinking models; general capabilities such as instruction following, tool use, text generation, and human preference alignment have also improved; it likewise enhances 256K long-context understanding, expandable to 1 million tokens.
Usage: The materials provide multiple paths for running and deployment. Through TUsing Transformers requires recommending the latest version, and transformers>=4.51.0; Qwen3-Instruct-2507 only supports non-thinking mode, its output does not contain a <think></think> block, and there is no need to specify enable_thinking=False; Qwen3-Thinking-2507 only supports thinking mode, and the default chat template automatically includes <think>, so the model output may only show </think> without an explicit opening tag. By default, Qwen3 models think before replying, and enable_thinking=False can be used to strictly control not generating thinking content; you can also use the /think and /no_think instructions in the system or user messages, and in multi-turn conversations the latest instruction is followed. For local running, you can use llama.cpp (llama.cpp>=b5401 is recommended for full Qwen3 support), Ollama (v0.9.0 or higher is recommended), LM Studio; it can also be run through ExecuTorch, MNN, mlx-lm (Apple Silicon, mlx-lm>=0.24.0), OpenVINO (Intel CPU or GPU), and other methods. For deployment, SGLang (sglang>=0.4.6.post1), vLLM (vllm>=0.9.0 is recommended), TensorRT-LLM (tensorrt_llm>=0.20.0rc3 is recommended) are supported, and on Ascend NPU you can access Modelers to search for Qwen3.
Tools and training: The material suggests that for tool use capabilities you can look at Qwen-Agent, which provides wrappers around the relevant APIs, supports tool use or function calling, and supports MCP; tool use can also be done through SGLang, vLLM, Transformers, llama.cpp, Ollama, etc. For training, it is recommended to use frameworks such as Axolotl, UnSloth, Swift, and Llama-Factory for fine-tuning such as SFT, DPO, and GRPO.
License and limitations: All open-weight models use the Apache 2.0 license, and the license files are located in their respective Hugging Face repositories. The material mentions that llama.cpp and Ollama use "rotating context management", which may affect long context; Ollama defaults to num_ctx 2048 and num_predict -1, and users are advised to set them correctly. In addition, SGLang and vLLM discard all reasoning_content fields when preprocessing API requests, which may affect the quality of multi-step tool use by Qwen3 thinking models; it is recommended to pass the content as-is without extracting the thinking content, and let the chat template handle it correctly. Ollama's naming may be inconsistent with Qwen's original naming; for example, qwen3:30b-a3b in August 2025 points to qwen3:30b-a3b-thinking-2507-q4_K_M. The material also mentions that users in mainland China can prioritize using ModelScope, and provides the VLLM_USE_MODELSCOPE=true and SGLANG_USE_MODELSCOPE=true environment variables.
