Alibaba Group has launched an upgraded version of its Qwen3 large language model (LLM) on July 22, 2025. The new model reportedly performs better in mathematics and coding evaluations than some competing models.

The Qwen3-235B-A22B-Instruct-2507-FP8 model shows improvements in instruction following, reasoning, text comprehension, and tool usage.

It scored 70.3 on the 2025 American Invitational Mathematics Examination, exceeding scores of 46.6 from DeepSeek-V3-0324 and 26.7 from OpenAI’s GPT-4o-0327.

In coding, it achieved 87.9 points on the MultiPL-E benchmark, slightly outperforming DeepSeek and OpenAI models, which scored 82.2 and 82.7, respectively.

However, it scored lower than Anthropic’s Claude Opus 4 Non-thinking, which received 88.5.

The upgrade from the previous Qwen3-235B-A22B-FP8 model includes an eightfold increase in content processing capacity. This new model accommodates up to 256,000 tokens per interaction.

Despite this advancement, it operates in a “non-thinking” mode, generating outputs directly without reasoning steps.

Additionally, Alibaba announced a collaboration to integrate a Qwen model with 3 billion parameters into HP’s “Xiaowei Hui” smart assistant for personal computers in China. This partnership aims to improve features such as document drafting and meeting summarization.

Recent Alibaba developments