After Deleting 10TB of Old Models, Regret: DeepSeek V3.2's World Knowledge Still Amazes
segmond · reddit · 2026-08-25
The author has deleted roughly 10TB of older models to save storage. While deciding whether to remove DeepSeek V3.2, they ran it one last time—and it answered their hardware question the best of all: absolute detail, familiarity, and accurate specs. Their takeaway: newer models are definitely smarter, more agentic and better at tool calling, but some older models remain remarkably clear and comprehensive in world knowledge. They're now considering re-downloading DS-0324 and K2 for prose/writing, and warn others not to blindly delete old models—some are still worth their weight.
More from Models
- GLM-5.3 weights will be released tomorrow — serige · 2026-08-27
- TokenSpeed adds Day-0 support for Qwen 3.8 Flash Next architecture — Alibaba_Qwen · 2026-08-27
- AI models show more creativity when talking to each other than in assistant persona — nabeelqu · 2026-08-27
- Claude Code shrinks by 45% in next version — jarredsumner · 2026-08-27
- OpenRouter leaderboard: Real token consumption data outweighs media hype — sujingshen · 2026-08-27
- Qwen 3.8-Next Released with Detailed Technical Report on Architecture — nrehiew_ · 2026-08-27