树莓派5跑Ollama：不到100美元跑本地LLM

树莓派5便宜又省电——但能跑现代LLM吗？选对模型就行。

Ollama on Raspberry Pi 5 2026: Run a Local LLM for Under $100

The Raspberry Pi 5 (8GB, ~$80) cannot run a 70B model - but it can run small models fine for chat, summarization, and light agent tasks. With Ollama, setup takes 15 minutes. Here is the honest performance picture.

Why the Pi 5 Is Viable

• 8GB RAM variant handles 3B-8B quantized models
• ~5-10 tokens/s on 3B models - usable for chat
• 5W power draw - runs 24/7 for pennies
• Official Raspberry Pi AI Kit adds 26 TOPS NPU for image models

Setup

# 1. Install Ollama (official script supports ARM64)
curl -fsSL https://ollama.com/install.sh | sh

# 2. Pull a small model
ollama pull qwen2.5:3b

# 3. Chat
ollama run qwen2.5:3b

Model Recommendations by RAM

Real-World Uses That Work

• Home assistant with a custom system prompt (8GB model, decent answers)
• Offline document summarization (3B is enough for short docs)
• Privacy-sensitive chat (data never leaves your network)
• Teaching/learning LLM concepts (inspect prompts and outputs freely)

What Does NOT Work

• Coding agents and complex reasoning (models too small)
• Large context windows (8GB RAM limits context to ~8-16K tokens)
• Real-time translation of long text
• Anything requiring a 30B+ model

FAQ

Is it worth it vs cloud APIs? For learning and privacy, yes. For heavy production workloads, a cloud API or GPU machine is faster per dollar.

Can I cluster multiple Pis? Yes - llama.cpp supports distributed inference, but network overhead limits gains; 2 Pis is practical, 4+ rarely pays off.

Does the AI Kit help with LLMs? The NPU accelerates vision/embedding models; text LLMs still run on the CPU. Speedup varies by model.