Ryzen AI Halo 192GB: Can It Run 300B Models on Your Desk?

Ryzen AI Halo just got a serious memory upgrade. On October 12, AMD introduced the next version of its Ryzen AI Halo developer platform, now built around the Ryzen AI Max+ PRO 495 processor with up to 192GB of LPDDR5x unified memory. AMD says that is enough to run 300-billion-parameter models locally at 4-bit quantization, and the Linux configuration is already on sale at Micro Center for $7,299.99.

• Processor: AMD Ryzen AI Max+ PRO 495 with 16 Zen 5 cores, 40 RDNA 3.5 GPU compute units and an NPU rated at up to 55 TOPS, per AMD.
• Memory: up to 192GB of LPDDR5x unified memory, with up to 160GB assignable to the GPU as VRAM; Borncity reports 8,533 MT/s on a 256-bit bus for 273 GB/s.
• Model size: AMD says the chip can run 300-billion-parameter and larger models at 4-bit quantization.
• Price and availability: $7,299.99 for the Linux system at Micro Center, with a Windows version using "Project Zenith" coming soon.

What Changed From the First Ryzen AI Halo?

The original Ryzen AI Halo developer platform arrived in June with the Ryzen AI Max+ 395. The new model moves to the PRO 495 and, more importantly, raises the memory ceiling to 192GB. For local AI work, memory capacity decides which models fit at all, so this is the upgrade that matters most.

Borncity lists the rest of the spec sheet: a 3.1 GHz base clock boosting to 5.2 GHz, a 2TB self-encrypting PCIe 4.0 SSD, two 40 Gbps USB4 ports, HDMI 2.1b, 10 Gigabit Ethernet, Wi-Fi 7 and Bluetooth 5.4. The processor's TDP is configurable from 45 W to 120 W, with 55 W as the default.

How Big a Model Can Ryzen AI Halo Run?

AMD's footnote is specific: with up to 160GB of dedicated graphics memory, the Ryzen AI Max+ PRO 495 is capable of running 300-billion-parameter and larger models at 4-bit quantization. AMD also highlights running several large agents at the same time, each with its own KV cache, because the CPU, GPU and NPU all share one memory pool.

That is where unified memory shines. A discrete graphics card tops out at its onboard VRAM, while a Strix Halo-class system can hand most of its RAM to the GPU. If you are sizing your own build, our explainer on how much RAM a local LLM needs maps model sizes to memory by quantization level.

Capacity Versus Speed

Expect the win here to be capacity rather than raw throughput. With 273 GB/s of memory bandwidth, token generation speed sits in the same class as other large unified-memory machines, so the Halo's appeal is fitting bigger models and more concurrent agents on one quiet desktop rather than topping speed charts.

Who Is Ryzen AI Halo For?

AMD pitches the platform at AI developers and creators who want to build and test locally, switching to the cloud only when a job calls for it. Borncity adds healthcare, finance and government teams that need to keep data on-premises. The Linux build ships with AMD's ROCm stack for PyTorch and TensorFlow, and the Windows version should widen the audience once it lands.

How Does It Compare to Other 192GB Mini PCs?

The 495 has been turning up in several compact systems this autumn, including the GMKtec EVO-X5 Pro with 192GB. AMD's own Halo stands apart as the reference design with a curated software experience and AMD's developer support behind it. Its premium price reflects that positioning. According to Borncity, Micro Center is selling it in stores only for now. For more compact AI hardware, browse our mini PC coverage.

Sources: AMD: Meet the Next AMD Ryzen AI Halo — October 12, 2026; Borncity — October 14, 2026.