[Official FAQ] Share your experience or follow-up questions below.

There is no single best model. These are general impressions only, and new versions come out constantly, so always test on your own tasks.

• Qwen: strong multilingual and Chinese support, a wide range of sizes, and specialized variants for coding and vision. A popular all-rounder.
• Llama: the largest ecosystem, lots of tutorials and fine-tunes, strong English. Read the license terms before commercial use.
• DeepSeek: the full flagship models are very large MoE models that most home machines cannot run. Locally, people usually run the distilled smaller versions, which are built on other base models. They tend to be good at reasoning but produce long "thinking" output.
• Mistral: efficient models with good English and European language support. Some releases use permissive licenses.
• Gemma: compact and efficient, a good fit for limited VRAM. Check the license terms.

A practical way to choose

• Pick two or three candidates that fit your memory.
• Write 10–20 real prompts from your own work.
• Run them with the same settings and compare answers side by side.
• Prefer "Instruct" versions for chat, and "Coder" versions for programming help.

Also check the license, supported languages, context length, and whether the model supports tool calling if you need it.

What are you using the model for? Tell us your task and hardware and the community can suggest candidates.