Kimi K3 Requires Full Message History: Most Users Are Using It Wrong
NielsRogge · x · 2026-08-04
Hugging Face member NielsRogge points out that Kimi K3 was trained in preserved thinking history mode, requiring the complete assistant message (including reasoningcontent and toolcalls) to be passed back as-is for multi-turn conversations and tool calls. He jokes that 99% of people are using Kimi K3 incorrectly.
More from Models
- Inside Anthropic's Prompt Engineering: 3-Layer Architecture and Task Decomposition — TansuYegen · 2026-08-04
- Joking About LLM Guardrails: Dev Proposes a 'Cuss Word Interceptor' — wavefnx · 2026-08-04
- MiniMax H3 Voice Model Glitch: Perfect Spanish Accent, Total Nonsense — Redditforgoit · 2026-08-04
- Why Do Chinese LLMs Outperform Western Labs in Frontend Development? — mouseofcatofschrodi · 2026-08-04
- DeepSeek Performs 'Just Fine' for 95% of Daily Tasks, Even Beating Sonnet — yacineMTB · 2026-08-04
- Ditching Claude Code: Migrating Complex Workflows to Open-Weight Models — TheZachMueller · 2026-08-04