Qwen
Also known as: Qwen models, Alibaba Qwen, Tongyi Qianwen
Qwen (short for Tongyi Qianwen, Alibaba's AI brand) emerged as one of the leading open-weight model families from China, competing in quality with Llama and Mistral for developers who want capable, freely downloadable models. The Qwen3 generation introduced hybrid thinking modes (a single model that can apply step-by-step reasoning or fast direct responses depending on the query) and strong multilingual performance, particularly in Chinese and English.
As of late 2025 through mid-2026, the Qwen3 lineup available in Amazon Bedrock and via Hugging Face includes Qwen3-Coder-480B-A35B (480B total / 35B active MoE, optimized for repository-scale code analysis and agentic tasks), Qwen3-Coder-30B-A3B (30B total / 3B active, coding and instruction following), Qwen3-235B-A22B (general purpose reasoning MoE), and Qwen3-32B (dense model for resource-constrained or edge environments). In April 2026, Alibaba released the Qwen 3.6 generation focused on real-world agentic workflows.
Qwen models support up to 256K token context windows natively (1M with extrapolation methods), tool calling, MCP integration, and structured output. They are available managed through AWS Bedrock and Alibaba Cloud, or as downloadable weights for self-hosting. For builders with data residency or sovereignty requirements, Qwen's Apache 2.0 license and self-hosting support make it an option worth evaluating, though geopolitical considerations around Chinese-origin models remain a factor in some regulated enterprise contexts.