Local LLM Performance Benchmarks: August 2026 Update
24
monthly benchmark update for local models. some new entries are impressive
curious if others have had similar results
6 replies
6 Replies
Join the discussion.
Log In to Reply
20
1
gpu prices hurt less if you go used 3090s. ive been running a pair of them for months and the perf per dollar is still hard to beat for anything under 30b params.
1
not sure benchmark scores tell the whole story. throughput on a shared machine vs dedicated inference hardware is such a different experience. what specs are these numbers based on?
2
-1
lm studio is convenient but does it let you inspect what data leaves the machine? asking genuinely - some of the telemetry defaults in these GUI tools are not obvious.
hmm the llama 3.3 70b numbers are interesting but i want to see how they hold up on multi-turn reasoning, not just single-shot benchmarks. those can be pretty misleading.