Just now, Qwen3 received a surprising upgrade!

The Tongyi Qianwen updated its flagship model Qwen3, releasing an updated version of the non-thinking mode (Non-thinking) called:

Qwen3-235B-A22B-Instruct-2507-FP8.

The new Qwen3 model has significantly improved general capabilities, including instruction following, logical reasoning, text understanding, mathematics, science, programming, and tool usage.

It performed excellently in many evaluations such as GQPA (knowledge), AIME25 (mathematics), LiveCodeBench (programming), Arena-Hard (human preference alignment), and BFCL (Agent capabilities), surpassing top open-source models like Kimi-K2 and DeepSeek-V3, as well as leading closed-source models like Claude-Opus4-Non-thinking.

In addition, the newly updated Qwen3 model has enhanced the following key performances:

· The model has made significant progress in long-tail knowledge coverage across multiple languages.

· In subjective and open-ended tasks, the model has significantly enhanced its ability to align with user preferences, providing more useful responses and generating higher-quality texts.

· Long text has been increased to 256K, and the context comprehension capability has been further enhanced.

To achieve optimal performance, Alibaba has also prepared recommended configurations for everyone:

· Sampling parameters: Temperature=0.7, TopP=0.8, TopK=20, MinP=0 · Appropriate output length: Output length is 16,384 Tokens, which is sufficient for instruction models.

· Standardized output format:

Math problems: Add "Please reason step by step and put the final answer in \boxed{}." in the prompt.

Multiple-choice questions: Add the following JSON structure in the prompt to standardize the response format: "Please only fill in the option letter in the answer field, for example: 'answer': 'C'."

Note: This model only supports non-thinking mode, and it will not generate blocks in the output. Also, there is no need to specify enable_thinking=False.

Currently, the new Qwen3 model is open-sourced on the ModelScope community and Hugging Face:

https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8

Original article: https://www.toutiao.com/article/7529701155111322162/

Statement: This article represents the views of the author. Welcome to express your opinion by clicking the [Up/Down] buttons below.