Leaderboard

Measured local LLM scores, per language and per GPU

Accuracy on real exams and decode speed, from our test PC and from people who shared a localllm report. Pick a language.

Language Measured by

Knowledge (translated): Global-MMLU-Lite, the same 400 questions in every language, so languages compare like for like. Local exams: INCLUDE, written in each country (ThaiExam for Thai). Math: MGSM. Translation: FLORES-101, chrF++ (0-100). Zero-shot, thinking off. Margins are about ±4 points (translated) and ±5 (local exams). Speeds are llama-server decode on the listed card. Raw data: data/leaderboard.json.

By language and by GPU

The same numbers as plain pages, one per language and one per graphics card.

English Arabic Chinese Hindi Japanese Spanish Thai

AMD Radeon RX 9070 XT

Where it is tested

One test PC is ours: Windows 11 with an RX 9070 XT. The rest of this table fills in from shared reports.

	AMD	NVIDIA	Intel	Apple
Windows	tested by us	untested	untested	-
Linux	untested	untested	untested	-
macOS	-	-	-	untested

Add your GPU

Ten to thirty minutes, no coding:

share a result

$ localllm tune
$ localllm eval --langs en,de     # your language instead of de
$ localllm report

report opens a prefilled GitHub issue with your GPU, RAM, OS, llama.cpp build, tuned settings and scores. No names, paths or keys; you read it before you submit it. How reports are checked.