Back to devhandbook.io
Free dev tool

Local LLM Runtime Picker

Ollama, llama.cpp, LM Studio, vLLM, Lemonade, or Jan — answer three questions and get the right runtime for your hardware and use case.

🍏 Apple Silicon (M-series)
🎮 NVIDIA GPU (CUDA)
🔴 AMD GPU (ROCm)
🧠 CPU only / low-power box
💬 Chat / assistant
⌨️ Coding (IDE integration)
🌐 API serving (OpenAI-compatible)
⚡ Max tokens/sec / batch
🛟 Ease of use
🚀 Max performance
🔒 Privacy / fully offline
🛠️ Max control / tinkering
How this works. The picker scores six runtimes against your three answers using a simple weighted matrix. It's a starting heuristic, not gospel — your exact GPU, model size, and tolerance for setup friction will shift the answer. For a deep dive, see the comparison table at the bottom.

Runtime comparison (2026 landscape)

Runtime Best for Weak at Vibe
OllamaDefault for most people; huge model libraryAMD/ROCm lag; less control; not max throughputBatteries-included
LM StudioGUI, easy model download, good Apple MetalNot for high-throughput servingPoint-and-click
llama.cppRuns on anything (CPU/NPU/edge); max controlCLI friction; no batteries includedThe engine underneath
vLLMProduction serving, high throughput, NVIDIASetup complexity; weak Apple supportServer-grade
LemonadeAMD GPUs/NPUs (official ROCm server)NVIDIA/Apple not the targetAMD's answer
JanLocal-first GUI, fully offline, no cloudYounger ecosystem; fewer modelsPrivacy-first