Start small, then measure
RAM is your computer's general working memory. Dedicated VRAM belongs to a graphics card; Apple Silicon shares one unified memory pool. Do not add the two numbers together as though they were interchangeable. Model weights, context, runtime buffers and other apps all need room.
A rough starting range
- Under 8 GB: do not assume comfortable local generation. Verify a tiny model or consider a supported cloud route.
- 8–15 GB: experiment with a 1–3B quantized text model and short context.
- 16–31 GB: begin with a 3–8B quantized model.
- 32 GB or more: begin with 7–14B rather than the largest download available.
These are conservative editorial starting points, not vendor-certified requirements. A model's exact quantization, context, architecture and runtime decide whether it fits. A smaller model may be faster but less capable for your task.
Check before downloading or buying
- Check OS/CPU/GPU support in the vendor requirements. LM Studio recommends 16 GB RAM and, on Windows, 4 GB dedicated VRAM; its x64 Windows build needs AVX2. Its Mac version supports Apple Silicon rather than Intel Macs.
- Choose one exact model file and a short test context. Larger contexts and parallel requests need more memory.
- In LM Studio's CLI, list downloaded model keys and estimate the chosen configuration:
lms ls
lms load --estimate-only <model_key> --context-length 4096Replace <model_key> with the actual key from your local list. The estimate is not a speed or quality benchmark. With Ollama, ollama ps shows the loaded model's processor split and context.
Chat fit is not agent fit
Documents, long conversations and tool calls can need far more context. Hermes currently documents a 64k minimum context for tool-enabled agents. A small model that answers a greeting is not thereby a suitable agent model. Validate tool support, context and task quality separately.
Start with LM Studio · Start with Ollama · Troubleshoot memory