DeepSeek

General-purpose chat, coding, summarization, and instruction-following tasks

This is the everyday workhorse. If you just need a smart, fast model for normal tasks, start here. Performance is genuinely competitive with GPT-4o at a tiny fraction of the cost.

Complex reasoning, math problems, logic chains, and tasks where showing your work matters

DeepSeek's answer to OpenAI o1. It thinks before it answers, which makes it slower but significantly better at multi-step problems. Pick this when accuracy on hard problems matters more than speed.

Running capable reasoning models locally or on modest hardware

These are smaller versions distilled from R1. The 7B and 8B run on a decent consumer GPU. The 32B and 70B need more serious hardware. If you want R1-style reasoning without API costs or data privacy concerns, these are your best option.

Historical reference; largely superseded by V3

Don't bother unless you have a specific reason. V3 is better in almost every way.

Code generation, debugging, and technical documentation

A coding-specialized model that holds up well against GitHub Copilot and CodeLlama. Worth testing if code quality is your primary use case, though V3 has largely caught up.