They reported a rate of 150 tokens per second after switching, but did not explain what caused the difference.
The same post reports 70 tokens per second with DeepSeek V4.1 Flash at 2.25-bit quantization. The user listed an Unsloth 6-bit quantization, 160,000-token context, 256 GB of DDR4 memory and the three GPUs; the post does not provide independent verification or a controlled comparison.