What to compare
- processor and accelerator
- RAM and accelerator memory
- model digest and quantization
- context and concurrency
- runtime revision and measured throughput
Read the sources
MLPerf Inference ↗
Inference system results with workload, hardware and submission rules.
Read the official results and their applicable benchmark round and division. Compare matching scenarios and quality targets. A datacenter result does not prove performance on a laptop.
llama.cpp benchmark tool ↗
A local measurement procedure for prompt processing and token generation under explicit runtime settings.
Public tool documentation. Downloading a model or running a benchmark needs separate permission. Record hardware, model digest, quantization, backend, context, batch settings and software revision.
Ollama model library ↗
Model pages and available variants for an Ollama-compatible setup.
Browse the published model pages. Availability in a library is not proof of access through a particular account. Check the model's own licence and variant. A listed size is not a complete working-memory estimate.
When to look again
Before choosing hardware or deploying a model, and after model, context, concurrency or runtime changes.
Give your agent this question
Separate memory estimates from measured compatibility. Propose a bounded benchmark using the exact hardware, model variant, context and runtime; do not claim a fit from parameter count alone.