Qwen 3.8 launches with 7b and 14b models

Alibaba unveiled the Qwen 3.8 model family on July 19, 2026, announcing 7-billion-parameter and 14-billion-parameter variants, with weights to be released under Apache 2.0 by year-end [Twitter][Alibaba Cloud Blog]. The Qwen 3.8 series builds on the transformer architecture of its predecessor, Qwen 2, but expands the context window to 32k tokens and incorporates a 2-trillion-token training corpus spanning 30 languages [Alibaba Cloud Blog]. Both model sizes ship with a unified inference API compatible with existing Qwen 2 deployments, enabling drop-in upgrades for applications that rely on Alibaba's cloud-native AI stack. The announcement also includes a benchmark suite showing 12% lower latency on V100 GPUs compared with Qwen 2-7B at the same batch size. The 30-language training mix delivers a 9% improvement in BLEU scores on low-resource language pairs versus Qwen 2 [Alibaba Cloud Blog]. The 12% latency reduction translates to roughly $0.018 per 1M token inference on a standard V100 instance, according to Alibaba's pricing calculator. By releasing the weights under Apache 2.0, developers can fine-tune Qwen 3.8 on proprietary data without negotiating separate licenses, a step that mirrors the open-weight strategy of Meta's Llama 3 [Alibaba Cloud Blog].