Alibaba releases Qwen2
Five model sizes from 0.5B to 72B parameters, trained on 27 additional languages beyond English and Chinese, with the smaller sizes under Apache 2.0.
- Open weights & ecosystem
- Models & capabilities
- Minor
Alibaba’s Qwen team released Qwen2, the successor to the Qwen family it had begun open-sourcing the previous year. The release spanned five sizes — 0.5B, 1.5B, 7B and 72B dense models plus a 57B mixture-of-experts model with 14B active parameters — with the 7B and 72B variants supporting context windows up to 128,000 tokens.
Training data expanded coverage to 27 languages beyond English and Chinese, including German, French, Spanish, Japanese, Korean and Arabic, addressing a common complaint about earlier open Chinese models being weak outside their two primary languages. Licensing was mixed: the 0.5B through 57B models were released under Apache 2.0, while the flagship 72B model and its instruction-tuned variant kept Alibaba’s more restrictive Qianwen licence. Alibaba reported state-of-the-art results among open models on a broad set of benchmarks, with particular gains in coding and mathematics over the prior Qwen1.5 generation.
Qwen2 arrived amid intense competition among Chinese labs — including DeepSeek and others — to lead open-weight benchmarks, a contest that continued through subsequent Qwen releases.