AMD Strix Halo: $1,499 Mini PC Runs 120B AI Models Locally

AMD's Strix Halo mini PC brings local AI inference to a new price point, offering 128 GB of unified memory in a $1,499 system that can run models up to 120 billion parameters without cloud dependency. For AI builders evaluating on-device inference, the trade-offs are clear: cost efficiency and large model capacity versus bandwidth constraints and a maturing software stack.

What happened

AMD released the Strix Halo, a mini PC built around the Ryzen AI Max Plus chip (codenamed Strix Halo). It uses a unified memory architecture that supports up to 128 GB of memory, allowing it to load and run large AI models with up to 120 billion parameters locally. The device delivers 34 tokens per second on a 120B model, trailing NVIDIA's DGX Spark by about 13% but at a $3,000 lower cost.

Real-world memory bandwidth is 122 GB/s, significantly below the advertised 256 GB/s. Entry-level systems start at $1,499, with a price per gigabyte of memory at $25.77 compared to Apple's M3 Ultra at $41.66 per GB. AMD also announced a next-generation Gorgon Halo chip planned for Q3 2026, featuring 192 GB of unified memory and support for up to 300 billion parameters.

Why AI builders should care

For teams building agent-heavy workflows or running local inference, the Strix Halo offers a practical alternative to cloud-based solutions. The unified memory architecture removes the VRAM ceiling that limits discrete GPUs, enabling models that physically don't fit on any single consumer GPU. This reduces operational costs and gives operators greater control over data for sensitive or latency-sensitive applications.

The device is particularly well-suited for workloads that prioritize cost efficiency over peak speed, such as batch inference, local RAG pipelines, and multi-agent systems where model size matters more than token generation rate.

Practical implications

When comparing Strix Halo to alternatives, the choice depends on workload characteristics. For generation-heavy tasks on large models, Strix Halo's 34 tokens/s is competitive with the DGX Spark at a fraction of the cost. However, NVIDIA retains a significant lead in prefill speed, particularly for long-context tasks where it is five times faster. This makes NVIDIA a better choice for document-heavy applications requiring rapid prefill rates.

For bandwidth-intensive workloads, Apple's M3 Ultra delivers 819 GB/s of memory bandwidth, far exceeding Strix Halo's real-world 122 GB/s. This gap may steer users toward Apple for tasks like large-context window processing or high-throughput inference.

AMD's ROCm software stack remains in preview with limited Windows compatibility, contrasting with NVIDIA's mature CUDA ecosystem. Vulkan provides a temporary workaround for some users, but the software maturity gap is a real consideration for teams needing immediate compatibility.

Caveats

• ROCm stack is still in preview and lacks full Windows compatibility, limiting software support breadth.
• Real-world memory bandwidth (122 GB/s) is far lower than Apple M3 Ultra's 819 GB/s, which may affect high-bandwidth workloads.
• Strix Halo may not match NVIDIA or Apple for peak speed or bandwidth in some workloads, particularly long-context prefill and bandwidth-intensive tasks.
• The advertised bandwidth of 256 GB/s is not achievable in practice, so builders should plan for the lower real-world figure.

FAQs

What is AMD Strix Halo and what does it do?

Strix Halo is an AMD mini PC built around the Ryzen AI Max Plus platform, designed for local AI inference with up to 128 GB of unified memory. It can run large models up to 120 billion parameters locally, reducing cloud dependency and giving users more data control for on-device workloads.

How many AI parameters can Strix Halo run locally?

The device is marketed to run models up to 120 billion parameters locally, thanks to its 128 GB unified memory architecture.

What memory and bandwidth does Strix Halo offer?

Strix Halo offers up to 128 GB of unified memory. Advertised memory bandwidth is 256 GB/s, but real-world measured bandwidth is 122 GB/s.

How does Strix Halo compare to NVIDIA DGX Spark and Apple M3 Ultra?

In tokens-per-second on a 120B model, Strix Halo trails the NVIDIA DGX Spark by about 13% (34 tokens/s vs higher for DGX Spark). Strix Halo has lower real-world bandwidth than Apple M3 Ultra (122 GB/s vs 819 GB/s) but offers a lower per-GB cost. DGX Spark may provide higher prefill speed for long-context tasks, while Strix Halo emphasizes cost efficiency and local inference.