← Back to feed

AMD shipped its first branded mini PC: the Ryzen AI Halo, built around the Ryzen AI Max+ 395 “Strix Halo” processor. It has attracted unusual hardware press coverage for a small-form-factor box because of what its memory spec implies for local AI inference.

The Number That Matters

128GB of LPDDR5X in a unified memory pool is the headline spec. For AI inference, memory is the binding constraint: model weights have to fit before you can run them. What 128GB enables:

• Llama 4 Maverick (17B active, ~400B total MoE): runs comfortably at full active-weight precision
• Any 70B dense model (Llama 4 Scout, Qwen-class 72B): fits with room for context
• Quantised 200B+ models at Q4: feasible
• GLM-5.2 at 2-bit (requires ~239GB): not yet — but one generation away on this platform

For context, a developer renting an 8xH100 instance gets 640GB of GPU memory — the ceiling for the same model class — at roughly $20-40/hour depending on provider and availability. The Ryzen AI Halo is priced around $2,999.

What the Chip Is

The Ryzen AI Max+ 395 is an NPU-augmented APU: CPU cores, AMD Radeon RDNA 3.5 integrated graphics, and a Ryzen AI NPU on a single die, all sharing the unified memory pool. There is no discrete GPU and therefore no separate VRAM ceiling. The shared pool is what makes the 128GB configuration meaningful — not just for system memory, but directly available to inference workloads.

The NPU handles int4/int8 quantised operations efficiently, which is the standard path for running 70B models at acceptable speed on edge hardware. For higher-precision inference or very large model evaluation, the GPU component does the heavy lifting.

Phoronix reviewed the device and found fully open-source software: AMD’s ROCm stack supports the chip, llama.cpp and Ollama run natively without proprietary drivers, and the device runs standard Ubuntu without modification. This is meaningfully different from Apple Silicon, where the open-source inference toolchain is still maturing and ROCm/CUDA-compatible workflows require bridging layers.

Why Now

Two dynamics have converged. First, frontier model weights at 4-bit quantisation now preserve most of the quality difference between compressed and full-precision versions. The memory ceiling matters more than raw compute for most developer use cases — inference throughput on a 70B Q4 model on the Halo is not datacenter-fast, but it is practical for single-developer workflows.

Second, tooling has simplified to the point where running a quantised 70B model locally is a fifteen-minute setup. LM Studio, Ollama, and llama.cpp handle model download, quantisation selection, and serving configuration without custom code.

The Ryzen AI Halo is not the first device in this territory. Apple Silicon M-series machines with 192GB configurations have been available since 2024 and have been used for local inference by developers on macOS. The Halo’s differentiation is the Linux-native ROCm stack: for workloads that require custom kernel integration, CUDA/ROCm-compatible tooling, or privacy-sensitive enterprise data that cannot leave a Linux environment, AMD now has a competitive entry in this hardware class.

What It Does Not Replace

Cloud inference for burst workloads, multi-user serving, and models that exceed local memory remains the dominant path. The Halo cannot run GPT-5.6 Sol, Claude Fable 5, or the 1T+ parameter open-weight models at competitive serving speed. For single-developer prototyping, local RAG, fine-tune evaluation, and privacy-constrained inference, it lands as AMD’s first serious entry in a market that Apple has owned by default.

The significance is structural: AMD, with ROCm and open-source tooling, entering local AI inference hardware at a competitive price point means the category now has two serious vendors. That’s a precondition for the tooling ecosystem maturing on Linux in a way it has not yet.