Why is my local LLM so slow?

Sounds like this?

• Words appear one at a time, slower than you can read
• A long pause before anything appears
• Your CPU fans spin up but the GPU stays idle

1. Make sure it runs fully on the GPU

If even part of a model spills into system RAM, it slows down a lot. In Ollama, run ollama ps: the PROCESSOR column should say 100% GPU. Anything like "40%/60% CPU/GPU" means the model is too big; pick a smaller one or a smaller quantization. In LM Studio, set GPU Offload to the maximum when loading. If it never uses the GPU, see model not using my GPU.

2. It may be thinking, not stuck

Reasoning models such as Qwen3, DeepSeek R1 and gpt-oss work through the problem before answering, which can take a while. That is normal. For quick chats you can turn it off for Qwen3: type /set nothink in Ollama, or add /no_think to your message. In LM Studio, look for the thinking or reasoning-effort setting next to the chat box.

3. Try a mixture-of-experts model

Models such as Qwen3 30B-A3B and gpt-oss 20B only use a small part of themselves for each word, so they run several times faster than ordinary models of the same size.

4. Plug in your laptop

On battery, laptops slow their GPU down. Plug in and choose the Best performance power mode in Windows settings.

5. The first reply is always slower

The model has to load from disk into memory first. Replies after that are quicker. Ollama unloads idle models after five minutes, so a pause between chats can mean another load.

6. Start a new chat

Very long conversations get slower, because the model rereads more text before each reply. Starting a fresh chat brings the speed back.