Siim Tolk

Live local LLM benchmarks

  • The test. 291 graded runs from what my assistant really does, suite version 2026-10-18. How the suite works.
  • The hardware. Two RTX 5060 Ti (16 GB each), running the same two-conversation setup as my everyday assistant. Last update 2026-10-26 09:24 UTC.

Quality against speed

Graded runs passed (Estonian and the uncensored set left out) against median prose generation speed of a single request. The vertical axis does not start at zero. Hover a point for its numbers; click it to jump to its row in the table below.

All results

Click a column to sort; click a setup name to open its weights on Hugging Face. Graded = passed/runs without Estonian and uncensored; thinking off = the same suite with reasoning disabled; follow-up = mean time to first token of the 2nd and later turns in two alternating 13k-token chats; 2 at once = total tok/s with two simultaneous requests. A dash means that test was not run for that setup.