[GH-ISSUE #696] Regarding the issue of 'max_tokens' #568

New issue

Closed

opened 2026-10-26 21:23:19 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 21:23:19 +00:00

Owner

Originally created by @qcjySONG on GitHub (Jun 21, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/696

I'd like to ask everyone: For the current 0528 version of R1, is the length control mechanism similar to QWEN3's built-in length control, or does it simply truncate when exceeding the limit (regardless of whether an answer has been output or whether it has exited the tag)?

So far in my testing, I haven't encountered any truncation. At the same time, it's hard for me to determine whether the shorter output is due to the parameters I set or simply because the model took a shorter COT path this time, leading to faster output. Hence, I'm here to ask for clarification!

请教一下大家，目前0528版R1的长度控制是类似QWEN3那种模型自身有某种长度控制机制，还是超出则直接截断（无论有没有输出答案，无论有没有离开）
我目前的测试还没有遇到截断的，同时我也很难说是我设置的参数导致模型的输出变短了，还是这一次模型走了较短的COT路径导致输出更快，因此来提问！

Originally created by @qcjySONG on GitHub (Jun 21, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/696 I'd like to ask everyone: For the current 0528 version of R1, is the length control mechanism similar to QWEN3's built-in length control, or does it simply truncate when exceeding the limit (regardless of whether an answer has been output or whether it has exited the </think> tag)? So far in my testing, I haven't encountered any truncation. At the same time, it's hard for me to determine whether the shorter output is due to the parameters I set or simply because the model took a shorter COT path this time, leading to faster output. Hence, I'm here to ask for clarification! 请教一下大家，目前0528版R1的长度控制是类似QWEN3那种模型自身有某种长度控制机制，还是超出则直接截断（无论有没有输出答案，无论有没有离开</think>）我目前的测试还没有遇到截断的，同时我也很难说是我设置的参数导致模型的输出变短了，还是这一次模型走了较短的COT路径导致输出更快，因此来提问！ ![Image](https://github.com/user-attachments/assets/df1b60c6-e1d4-4108-96c6-9f339712463e)

bsakaguchi closed this issue

2026-10-26 21:23:19 +00:00

bsakaguchi commented

2026-10-26 21:23:20 +00:00

Author

Owner

@qcjySONG commented on GitHub (Jun 21, 2025):

好的，经过我的尝试是超出Max_token则直接截断，非常期待未来的R1或者R2等版本有对应的思维链长度控制能力！

@qcjySONG commented on GitHub (Jun 21, 2025): 好的，经过我的尝试是超出Max_token则直接截断，非常期待未来的R1或者R2等版本有对应的思维链长度控制能力！