AMD Instinct MI250X~649 tok/s (Q4)

deepseek-ai/DeepSeek-OCR-2 speed on AMD Instinct MI250X

Quantization-specific throughput and VRAM requirements for deepseek-ai/DeepSeek-OCR-2 running on AMD Instinct MI250X.

Speed Snapshot

Topline estimate from compatibility data

Modeldeepseek-ai/DeepSeek-OCR-2

GPUAMD Instinct MI250X

Q4 speed649 tok/s

Q4 VRAM required2GB

Data Source

Calculation and benchmark status

Speed values come from the compatibility dataset (`estimatedTokensPerSec`) and are sorted by quantization.

For full verdict logic and alternate GPUs, see the canonical compatibility page.

Quantization Speed Table

Quantization	VRAM needed	VRAM available	Speed	Verdict
Q4	2GB	128GB	649 tok/s	✅ Fits
Q8	4GB	128GB	378 tok/s	✅ Fits
FP16	8GB	128GB	246 tok/s	✅ Fits