Your machine
Ollama and llama.cpp split layers across cards automatically. Pooling buys you capacity, not speed — throughput stays near single-card, so the estimate below is for one card.
What you'll use it for pick any
about 6,000 words of conversation before it forgets. Longer needs more memory, and this is the step most calculators leave out.
What matters more
How this is worked out
Real file sizes, not a rule of thumb
Every size here is the actual byte count of the GGUF files in the repository, read live from Hugging Face. Nothing is estimated from a parameter count.
The context window is part of the sum
A model's weights are only half the story. The KV cache grows linearly with context and is calculated from the model's real layer and attention-head counts.
Speed is a range, deliberately
Generation is bounded by memory bandwidth. We show a band rather than a single number because a single number would imply a precision that nobody can honestly offer.