DeepSeek V3.1 Plagued by 'Extreme' Token Insertion, Disrupting Code Output

DeepSeek V3.1, a recently released large language model by DeepSeek AI, is reportedly suffering from a significant bug that causes it to randomly insert the token "extreme" or its Chinese equivalent "极" into generated outputs. This issue, which surfaced almost simultaneously within both English and Chinese AI communities, was highlighted in a tweet by user "Rookie" on August 25, 2025. The unexpected token insertions are particularly problematic for structure-sensitive tasks like code generation, where even minor deviations can lead to broken or unusable results.

Users have provided concrete examples of the bug's disruptive nature. In one instance, an expected output of "time.Second" was corrupted to "time.Se极," while another saw "time.Se extreme" generated. These random insertions, which include the simplified Chinese character "极" (ID: 2577) and the traditional Chinese character "極" (ID: 16411), appear in unexpected places, often as a top or secondary choice in the model's token predictions, severely impacting the integrity of its responses.

The root cause of the "extreme" token bug is currently under community speculation, with leading theories pointing to polluted training data. Discussions suggest that the model may have inadvertently learned to use "极" as a special symbol, potentially amplified through subsequent reinforcement learning and self-distillation processes. Notably, similar anomalous token behaviors have been observed in earlier DeepSeek V3 versions (0324) and some Qwen3 models, indicating a recurring challenge in the development of large language models.

DeepSeek V3.1 itself, quietly launched on August 19, 2025, has otherwise been lauded for its advanced capabilities, particularly in programming and reasoning. It features a 685B-parameter hybrid reasoning architecture and boasts a 128k context length, offering significant cost advantages and strong performance in code generation and debugging. Its initial reception highlighted its potential as a strong contender in the open-source AI landscape, making the current bug a notable setback.

As of the latest reports, DeepSeek AI has not issued an official statement or announced a fix for the "extreme" token insertion bug. The AI community is actively discussing potential workarounds, including the use of logit bias or GBNF grammars to restrict unwanted tokens. The ongoing dialogue underscores the collaborative efforts within the open-source community to address and mitigate such unforeseen challenges in advanced AI models.