Qwen (Tongyi Qianwen)
Qwen (通义千问) is Alibaba Cloud's LLM family, with Chinese capability among the strongest of current open-source models. The latest Qwen2.5 series surpasses same-size Llama on MMLU, C-Eval, HumanEval and other benchmarks.
Version history
- Qwen 1.5: 0.5B / 1.8B / 4B / 7B / 14B / 72B full tier.
- Qwen 2: native 128k context, Chinese boost.
- Qwen 2.5: training data 7T → 18T tokens, big instruction-following improvement.
- Qwen3: released 2025, native thinking mode.
Derived series
- Qwen-VL: multimodal (image understanding).
- Qwen-Audio: audio understanding.
- Qwen-Code: code-specialized.
- Qwen2-Math / Qwen2.5-Math: math reasoning.
- QwQ (Qwen with Questions): reasoning-enhanced.
Chinese strengths
- Optimized Chinese tokenizer, far higher tokenization efficiency than Llama.
- High proportion of Chinese in training data, deep cultural context understanding.
- Long-time #1 on C-Eval Chinese leaderboard.
Best for
- Chinese scenarios: prefer Qwen.
- On-device deployment: 0.5B / 1.5B series.
- Local RAG: Qwen2.5 + text-embedding series work well together.