Back to catalog
␣extreme
The key token of DeepSeek V3.1's “极” bug (id 15075, with leading space): the model randomly injects “极”, “極” or “extreme” into generated text.
Raw token
" extreme"
␣marks a real space; the symbol itself is not part of the token, and copying preserves it.
Observed behaviors
- 01From 2025-08, many users reported V3.1 inexplicably emitting “极”, “極”, “极速” or “extreme” mid-generation, at random — e.g. `time.Second` becoming `time.Se极`, `V1` becoming `V极` (logprob -0.279, the top-1 candidate); Tencent Codebuddy would auto-complete “极速赛车开奖” and “极速电竞”.mp.weixin.qq.com
- 02Tokens involved: “ extreme” (id 15075), “极” (id 2577), “極” (id 16411) — “极” sits right next to “...” (2576) in the vocabulary; Fireworks' FP8 full-precision deployment reproduces it too, ruling out quantization.mp.weixin.qq.com
- 03The same symptom appears in Qwen3-235B-A22B-Instruct-2507 and Qwen3-Coder-30B-A3B — the report speculates the two labs may have used the same polluted data; GLM 4.5 is unaffected.mp.weixin.qq.com
- Observed or documented on
- DeepSeek