DeepSeek local quantized version exhibits weird CoT: normal output but caveman speech
sk1kn1ght · reddit · 2026-08-04
A developer running a local q8 quantized version of DeepSeek V3 (0731) via unsloth and llama.cpp noticed a bizarre reasoning behavior: while the final text output is perfectly normal, the model's internal chain-of-thought (thought text) is highly unusual. The CoT uses the pronoun "We" and breaks down into what the user describes as "caveman speech." The poster is asking the community if this is an expected model trait or a misconfiguration.
More from Models
- Tipping Your LLM? Testing Claude's Safety and Alignment Logic — hrishioa · 2026-08-04
- OpenMed Plans to Fine-Tune Qwen3 27B into the Best Local Medical AI Model — MaziyarPanahi · 2026-08-04
- Developer Reverse-Engineers and Rewrites Console Games Rapidly Using DeepSeek — yacineMTB · 2026-08-04
- Yacine Benchana Slams Anthropic's Guardrails: Actively Sabotaging LLM Development — yacineMTB · 2026-08-04
- DeepSeek Harness Praised for 'Unlimited Scalability' by Community — teortaxesTex · 2026-08-04
- SenseNova U1 Pro Unifies Multimodal Understanding and Generation at 8K — s_mohinii · 2026-08-04