Four Gaming Chips Run Huge AI Model at Home, Beating Cloud

A Reddit user with four AMD Radeon AI PRO R9700 graphics cards has successfully run a huge AI model called DeepSeek-V4-Flash-0731 at home. The model is so large that it normally requires expensive cloud servers. But by using a special engine called 'affinity,' the user got it working on just four consumer-grade cards. The setup generated text at 79 tokens per second, which is fast enough for real-time conversation. This is a big deal because it shows that powerful AI can run on hardware you can buy, not just in massive data centers.

The previous version of this engine refused to start on more than two cards because the developer only had two and simulated four, which produced garbage output. But when tried on four real cards, it worked perfectly. All 11,008 experts (the model's internal knowledge units) fit into the graphics cards' memory, so no data had to be streamed from the main computer memory. That's why it's so fast. The user also shared tips for others who want to try it, like using a specific placement strategy.

This matters because running AI locally means you don't have to pay monthly fees to cloud providers, and your data stays private. It also reduces reliance on internet connectivity. While the hardware is still expensive (four high-end graphics cards), it's a one-time cost that could pay off over time. Plus, it's a hobbyist achievement that pushes the boundaries of what's possible on consumer hardware. The user is offering to run more tests, so we might see even more improvements soon.