Back to the lesson

Local model planner

Choose your hardware and a model. The planner estimates how much memory the model needs, whether it fits, and how fast it will feel, then writes the Ollama commands to run it. The calculation happens in your browser and sends none of your choices anywhere. Like every page on this site, it counts page views with the Foundation's self-hosted metrics, which use no cookies.

Your hardware

Computer type
System memory: 16 GB
GPU video memory: 24 GB
How fast is your memory?

Your model

Model
Size class

Typical shapes for models of this class. Real models vary, so check the model card.

Quantization
Context window: 4,096 tokens
Context cache precision

How the estimate works

  • Weights = parameters × bits per weight ÷ 8. Olmo 3 7B at 4-bit comes to about 4.4 GB, close to the 4.5 GB download.
  • Context cache = 2 × layers × KV heads × head size × bytes × tokens. It grows with every token of context.
  • Overhead is about 1 GB for buffers.
  • Speed ≈ 60% of memory bandwidth ÷ bytes read per token. Mixture-of-experts models read only their active parameters.

These are teaching estimates. Models with sliding-window attention can use less cache, and real speed depends on the runtime and your system. Run ollama ps to see what actually happened.

Will it fit?

Ready
WeightsContext cacheOverheadFast memorySlower system RAM
Weights
Context cache
Overhead
Total needed
Fast memory available
System RAM available for spill

How fast will it feel?

0tokens per second

Olmo 3 on this machine

VariantNeedsResultSpeed

Commands for this setup

What ollama ps should show


        
Illustrative output. The PROCESSOR column tells you whether the model is on the GPU, the CPU, or split.

      

    

  

  
This is a teaching tool, not a benchmark. Model names, sizes and defaults change over time, so use it to learn the method and then check the current model card and the Ollama documentation. Dedicated to the public domain under CC0 1.0.