Qwen3 Launches Open-Weight AI Models Rivaling Top Industry Competitors
The Qwen team releases Qwen3, a new family of open-weight large language models that achieve top-tier benchmark results in coding and math. The release includes efficient Mixture-of-Experts models and smaller dense models available under the Apache 2.0 license.
The Qwen development team announces the release of Qwen3, the latest generation of their large language model family. The flagship model, Qwen3-235B-A22B, achieves highly competitive results against top industry models like DeepSeek-R1, o1, and Grok-3 in coding, math, and general capabilities. Smaller models in the family also show impressive efficiency, with the Qwen3-30B-A3B outperforming earlier models that use ten times as many activated parameters.
This release includes two Mixture-of-Experts (MoE) models and six dense models, all offered as open weights under the Apache 2.0 license. The MoE models feature a massive 235 billion total parameters with 22 billion activated, and a smaller 30 billion total parameter version with only 3 billion activated. The dense models range from 0.6 billion to 32 billion parameters, providing options for various hardware constraints and deployment needs.
Qwen3 models introduce a hybrid thinking approach that allows them to switch between quick responses and step-by-step reasoning modes depending on the complexity of the task. Users access these models easily through platforms like Hugging Face, ModelScope, and Kaggle, or deploy them locally using tools like Ollama and llama.cpp. The team makes the models available for interactive testing through the Qwen Chat web interface and mobile application.