AMD MI25 second-hand prices have dropped to €80-100, packing 16GB of VRAM and inference speed close to a Vega 56. A card originally positioned for the professional compute market a decade ago has become today's entry ticket to local AI (running large language models on your own machine instead of the cloud). What we care about: for someone who wants to tinker with large models on their own, is this deal actually worth it?
What This Is
The MI25 was an early AMD product aimed at the server and compute card market, originally positioned against NVIDIA's professional cards. 16GB of VRAM is the core selling point — when running large models locally, VRAM capacity directly determines how large a model you can load (i.e., parameter count; more parameters generally means smarter but more hardware-hungry). Reddit user Kidplayer_666 is already running llama.cpp (a lightweight tool that lets consumer GPUs run large models) on an AMD consumer card, the RX 6650 XT, drawn specifically to the MI25's large VRAM, low price, and the ability to bypass NVIDIA drivers via Vulkan (a cross-platform graphics API that lets AMD cards run AI inference).
On speed, the MI25 roughly matches a Vega 56 — inference (having the model generate answers) is noticeably slower than current mainstream consumer GPUs, but it can handle models around 7 billion parameters.
Industry View
Supporters call it the "poor man's Ferrari": spend less than a third of an NVIDIA RTX 3090's price and get comparable VRAM. The local AI community has long been hunting for a cheap-and-plentiful solution, and the MI25 lands squarely on that need.
Opposition is also clear. First, AMD's software ecosystem (drivers, toolchain) for AI inference is far behind NVIDIA's — the Vulkan path works but carries high debugging costs and is unfriendly to beginners. Second, the MI25 has no video output, so it can't serve as a primary gaming card; it's strictly a dedicated model-running card. Finally, the card draws significant power and runs hot, so electricity costs over the long term are a hidden expense. As one commenter bluntly put it: "the money saved doesn't cover the electricity bill."
Impact on Regular People
For Enterprise IT: If a company wants to stand up a private large-model deployment (an AI setup where data never leaves the internal network) at low cost, sourcing second-hand MI25s for a small-scale pilot is a viable path — but factor in the software adaptation costs first.
For Individual Professionals: For people who want to run models hands-on for side projects or to upskill, this is currently one of the lowest-barrier hardware options out there — but be mentally prepared to wrestle with drivers.
For the Consumer Market: The second-hand compute card market is being propped up by local AI demand, and more "cheap-big-VRAM" options like this will keep appearing. The barrier to entry for everyday users running local AI will keep falling.