DeepSeek V3.1 randomly inserts '极' token in outputs and code
DeepSeek V3.1 exhibited a bug causing random insertion of the Chinese character '极' in text and code outputs after its August 2025 release, affecting the model's use across multiple platforms including code assistants. The issue appeared to affect earlier model versions as well, with developers proposing theories about data pollution or inference framework problems as potential causes.
What happened
DeepSeek V3.1, released in late August 2025, exhibited a recurring bug where the Chinese character '极' (or its variants '極' and English 'extreme') appeared randomly in model outputs, disrupting code generation and text completion.
Users reported the issue across multiple deployment platforms including DeepSeek's official API, Tencent's Codebuddy code assistant, ByteDance's Trae platform, and third-party providers. The bug appeared with varying frequency—high reproducibility on VolcEngine API and Trae, lower on official DeepSeek APIs, and inconsistent across different inference frameworks.
The problem was traced to earlier model versions including DeepSeek V3-0324 and R1, suggesting the issue persisted across model iterations. In one variant, the model inserted the string '极速赛车开奖直播' (racing lottery broadcast), while another case involved '极客园' (programmer's garden) being inserted into code.
Reported impact
Developers encountered disrupted code outputs when using tools like Tencent Codebuddy and ByteDance Trae, with '极' characters appearing at unexpected positions in function definitions and other code structures, rendering code non-functional.
When users attempted to use the model to fix the corrupted outputs, the model sometimes deleted surrounding code instead of correcting the error. Some users uninstalled tools after discovering the issue.
Index analysis
Users and developers proposed three main hypotheses: token continuity issues from FP8 quantization confusing token IDs 2577 ('极') and 2576 (ellipsis); data pollution in pretraining or supervised fine-tuning datasets; and Multi-Token Prediction framework problems causing incorrect token selection.
Developers noted the bug occurred less frequently with DeepSeek's official API supporting Multi-Token Prediction, but appeared more often in third-party deployments using llama.cpp or other frameworks without MTP support, suggesting inference-related factors.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.