Best local LLM for Japanese

Measured Japanese scores of quantized local models, per GPU. Highest average so far: qwen3.8-27b-iq2 on AMD Radeon RX 9070 XT.

Model	GPU	Decode tok/s	Knowledge (translated) %	Local exams %	Math %	Translation chrF++	Source
`qwen3.8-27b-iq2`	AMD Radeon RX 9070 XT	40	65.8	77.9	-	-	maintainers
`qwen3.5-9b-q4`	AMD Radeon RX 9070 XT	42	68.0	82.3	81.6	44.8	maintainers
`qwen3.8-27b-q3`	AMD Radeon RX 9070 XT	50	73.5	87.6	-	46.4	maintainers
`gemma4-26b-a4b-qat`	AMD Radeon RX 9070 XT	90	74.5	81.9	-	48.0	maintainers
`qwen3.5-4b-q4`	AMD Radeon RX 9070 XT	78	65.5	70.7	76.8	43.7	maintainers
`gemma4-e4b-q4`	AMD Radeon RX 9070 XT	105	60.8	69.5	76.4	46.7	maintainers

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