Kimi K3 Still Heavily Uses English in Chinese Reasoning
emollick · x · 2026-07-19
Ethan Mollick observed: when asking Kimi K3 in Chinese to pick an unconventional poem that corresponds to an LLM, the model's generated chain-of-thought contained **95.5% English characters and 88% English words**, even though the prompt was explicitly aimed at Chinese readers discussing Chinese poetry. The attached image illustrates his reasoning process during retrieval and fact-checking. The key takeaway is that the model's internal thought trajectory still heavily relies on English when processing Chinese tasks. This post is more of an empirical observation of model behavior and language preferences.
Related event: Testing Kimi K3: 32-Page CoT Loops and English-Dominant Reasoning(5 posts)→
More from Models
- Grok Build CLI v0.2.108 adds cross-machine session resume and editor handoff — Scobleizer · 2026-07-21
- Gemini 3.6 Flash reportedly appears in Antigravity under `gemini-3.6-flash-tiered` — gaganghotra_ · 2026-07-21
- Grok 4.5 tops a long-horizon terminal benchmark, Elon Musk says — elonmusk · 2026-07-21
- K3 seems faster on the biggest coding plan than through OpenRouter, user says — xeophon · 2026-07-21
- Andrew Ng-style distillation joke turns model reuse into a theft punchline — pmddomingos · 2026-07-21
- Grok’s Twitter/X search quality appears to have regressed — ivan_bezdomny · 2026-07-21