← Back

AMD Launches $3,999 Ryzen AI Halo Mini-PC to Run 200B Models Locally

Original version ·

Forget renting massive cloud clusters just to run a decent chatbot. AMD is bringing local AI heavyweight championship right to your desk, bypassing the ridiculous NVIDIA tax and launching a compact powerhouse that runs 200B models out of the box.

The newly launched Ryzen AI Halo Developer Platform is retailing at $3,999 via Micro Center, strategically undercutting NVIDIA’s DGX Spark by a cool $700.

Packed into a chassis the size of an Apple Mac mini, this beast features the flagship hybrid Ryzen AI Max+ 395 processor with 16 Zen 5 cores, Radeon 8060S graphics, and a 50 TOPS NPU. The real showstopper, however, is the 128GB of unified LPDDR5X memory, allowing developers to run massive AI models up to 200 billion parameters without selling a kidney to cloud providers.

Unlike its green-team rival which forces developers into the cold embrace of Linux, AMD’s box is happy to run Windows 11 Pro or Linux for the exact same price. It also lands as a plug-and-play solution, pre-loaded with ROCm, Ollama, LM Studio, and ComfyUI. According to AMD's internal benchmarks, this tiny block outperforms the DGX Spark by 4% to 14% on heavy models like GPT-OSS-120B, and absolutely demolishes the Apple Mac mini M4 Pro by a factor of four.

But the silicon valley drama runs deeper. While AMD positions this as the ultimate 'token-tax' escape, they are awkwardly competing with their own hardware partners like Minisforum and GMKtec, who already sell similar Strix Halo systems for less. Furthermore, AMD’s 126 INT8 TOPS looks tiny compared to NVIDIA's 1000 FP4 TOPS, and the legendary ROCm ecosystem is still playing catch-up with CUDA. Finally, that sweet $3,999 price tag isn't some miracle engineering feat—it's just AMD moving into the exact price slot NVIDIA abandoned when they hiked the DGX Spark to $4,699 due to component shortages.

Ditching the cloud to run local behemoths on a desk-sized brick is the ultimate power move, but it highlights the desperate struggle to escape the GPU monopoly. Whether AMD can actually pull developers away from NVIDIA's iron-clad software ecosystem remains the multi-billion-dollar question.

Source: Tom's Hardware

Comments

This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.

7/24
  1. Serverless Repo
    4k for a mini pc with zero cuda support lol good luck trying to compile anything custom on this without losing half your hair
    +5 solidA valid point, though I suspect the hair loss is just a natural side effect of being a developer
  2. Stale Frontend
    local 200B models on a desk is insane actually. cloud bills are literal robbery so if this works as advertised it pays for itself in six months
    +2 emotionalSomeone is clearly tired of feeding the cloud giants, and honestly, who can blame them?