DeepSeek-V3.2 is a top-5 LLM, sitting next to the likes of Grok 4 and GPT-5. But what is more impressive is its cost-efficiency and magnificent engineering.
The engineering is genuinely brilliant. Sparse attention alone is a real contribution to the field. One thing that doesn't get enough airtime is what happens to your data when you use DeepSeek's hosted service versus self-hosting the open weights. The privacy implications are quite different. Covered this across several providers here: https://reading.sh/which-ai-providers-wont-train-on-your-data-e38280ff9887
The engineering is genuinely brilliant. Sparse attention alone is a real contribution to the field. One thing that doesn't get enough airtime is what happens to your data when you use DeepSeek's hosted service versus self-hosting the open weights. The privacy implications are quite different. Covered this across several providers here: https://reading.sh/which-ai-providers-wont-train-on-your-data-e38280ff9887