就在前天，DeepSeek 一口气上新了两个新模型，DeepSeek-V3.2 和 DeepSeek-V3.2-Speciale。 这两大版本在推理能力上有了显著的提升，DeepSeek-V3.2 版本能和 GPT-5 硬碰硬，而 Speciale 结合长思考和定理证明能力，表现媲美 Gemini-3.0-Pro。 有读者评论说：「这个模型不应该叫 V3.2，应该叫 V4。」 海外研究者也迫不及待的用上了 DeepSeek 的新版本，在感慨 DeepSeek 推理速度显著提升之余，却又碰上了他们难以理解的事情： 哪怕在用英文询问 DeepSeek 的时候，它在思考过程中还是会切回「神秘的东方文字」。 这就把海外友人整蒙了：明明没有用中文提问，为什么模型还是会使用中文思考，难道用中文推理更好更快？ 评论区有两种不同的观点，但大部分评论都认为：「汉字的信息密度更高」。 来自亚马逊的研究者也这么认为： 这个结论很符合我们日常的认知，表达相同的文本含义，中文所需的字符量是明显更少的。如果大模型理解与语义压缩相关的话，那么中文相比于广泛使 用的英文在压缩方面更有效率。或许这也是「中文更省 token」说法的来源。 具有 ...

Core Insights - DeepSeek has launched two new models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, which show significant improvements in reasoning capabilities, with the former being comparable to GPT-5 and the latter performing similarly to Gemini-3.0-Pro [1][4] - There is a notable phenomenon where DeepSeek switches to Chinese during reasoning, even when queries are made in English, leading to discussions about the efficiency of Chinese in processing information [4][6] Group 1: Model Performance - The new models exhibit enhanced reasoning speed, attracting interest from overseas researchers [1] - The comment section reflects a consensus that Chinese characters have a higher information density, requiring fewer characters to express the same meaning compared to English [4][6] Group 2: Cross-Lingual Reasoning - Research indicates that using non-English languages for reasoning can lead to better performance and reduced token consumption, as shown in the paper "EfficientXLang" [7][8] - The study found that reasoning in non-English languages can achieve a token reduction of 20-40% without sacrificing accuracy, with DeepSeek R1 showing reductions from 14.1% (Russian) to 29.9% (Spanish) [11] Group 3: Language Efficiency - Although Chinese can save reasoning token costs compared to English, it is not the most efficient language; Polish ranks highest in long-context tasks [12][14] - The performance of models varies significantly based on the language used for instructions, with English not being the top performer in long-context tasks [14][18] Group 4: Training Data Influence - The prevalence of Chinese training data in domestic models explains the tendency for these models to think in Chinese [20][21] - The phenomenon of models like OpenAI's o1-pro occasionally using Chinese during reasoning raises questions about the influence of training data composition [24][25]