Best local LLM for Hindi

Measured Hindi 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	69.5	71.0	-	62.8	maintainers
`qwen3.8-27b-q3`	AMD Radeon RX 9070 XT	50	69.0	74.3	-	58.1	maintainers
`gemma4-e4b-q4`	AMD Radeon RX 9070 XT	105	57.8	71.0	-	60.9	maintainers
`qwen3.5-9b-q4`	AMD Radeon RX 9070 XT	42	61.0	70.2	-	56.9	maintainers
`qwen3.5-4b-q4`	AMD Radeon RX 9070 XT	78	54.2	61.2	-	53.4	maintainers
`qwen3.8-27b-iq2`	AMD Radeon RX 9070 XT	40	56.2	55.5	-	-	maintainers

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