Qwen

Qwen (通义千问, Tongyi Qianwen) is Alibaba's foundation-model family, open-weight since 2023. Versions: Qwen 1, Qwen 2, Qwen 2.5, Qwen 3 (2025). Tier structure includes dense models from 0.5B to 72B and MoE variants scaling to 235B total parameters. Qwen 3 is the 2026 leading open-weight Chinese LLM · it also performs strongly on English and multilingual benchmarks, putting pressure on Llama in the open-source tier.

Qwen architecture: decoder-only transformer with RoPE, SwiGLU, similar to Llama. Qwen 3 235B is an MoE with 22B active parameters. Qwen's differentiator is training data · optimized for Chinese and multilingual coverage with heavy filtering for quality. Qwen ships under Apache 2.0 for most variants (truly permissive) · more open than Llama's custom license. The Qwen ecosystem includes Qwen-VL (multimodal), Qwen-Audio, Qwen-Coder (specialized), and many community fine-tunes.

Qwen 2.5 72B outperformed Llama 3.1 70B on many multilingual benchmarks. Qwen 3 235B MoE activates 22B per token, putting it between dense 70B and dense 405B in quality. Training: 18T+ tokens for Qwen 2.5. Alibaba uses their Ascend 910 chips alongside NVIDIA for training · one of the few non-NVIDIA-dependent Chinese training stacks. Qwen-Coder is SWE-bench-competitive among open-weight models. The Apache 2.0 license removes commercial friction entirely · contributing factor to Qwen's rapid adoption in 2025-2026.