Best local LLM for Arabic

Measured Arabic scores of quantized local models, per GPU. Highest average so far: gemma4-26b-a4b-qat on AMD Radeon RX 9070 XT.

Model	GPU	Decode tok/s	Knowledge (translated) %	Local exams %	Math %	Translation chrF++	Source
`gemma4-26b-a4b-qat`	AMD Radeon RX 9070 XT	90	71.5	73.6	-	61.9	maintainers
`qwen3.8-27b-q3`	AMD Radeon RX 9070 XT	50	70.8	71.2	-	59.0	maintainers
`qwen3.5-9b-q4`	AMD Radeon RX 9070 XT	42	65.8	66.4	-	57.5	maintainers
`gemma4-e4b-q4`	AMD Radeon RX 9070 XT	105	62.5	66.0	-	60.8	maintainers
`qwen3.5-4b-q4`	AMD Radeon RX 9070 XT	78	63.0	61.6	-	55.4	maintainers
`qwen3.8-27b-iq2`	AMD Radeon RX 9070 XT	40	60.8	57.2	-	-	maintainers

Zero-shot, thinking off; margins about ±4-5 points. How each column is measured: leaderboard. Missing your GPU? Run localllm eval --langs en,ar and share it with localllm report.