How fast does a local LLM run on your GPU or Mac?
1,127 measured speed results across 281 GPUs, Macs and APUs, each linked to its public source. No guesses in this table.
Leaderboard
Llama 2 7B Q4_0 is the standard llama.cpp test model, so it has the most comparable results. Generation speed (tg128) is what you feel when chatting.
Showing measured results only. How we collect and check them.
New models that fit your card, weekly
Every Monday: new open models grouped by how much VRAM they need, plus the newest measured benchmarks. Free, one email a week.
Sources in this release
- LocalScore (localscore.ai)
- llama.cpp Apple Silicon thread (#4167)
- llama.cpp CUDA scoreboard (#15013)
- llama.cpp ROCm scoreboard (#15021)
- llama.cpp Vulkan scoreboard (#10879)
- llama.cpp benches/ (maintainer runs)
- llama.cpp discussion #15396
- llama.cpp discussion #16578