AMD has unveiled the next version of its Ryzen AI Halo Developer Platform, a compact desktop built for running serious AI workloads locally. Powered by the new Ryzen AI Max+ PRO 495 processor and packing up to 192GB of unified memory, it is designed to run 300-billion-parameter language models entirely on your desk — no cloud required. Here is everything we know.
What Is the Ryzen AI Halo Developer Platform?
The Ryzen AI Halo is not a traditional desktop PC. It is a compact developer workstation — roughly the size of a small mini-PC — purpose-built for building, testing, and running AI models locally. The first version arrived earlier this year with the Ryzen AI Max+ 395 chip, 128GB of memory, and a $3,999 price tag, rated for models up to 200 billion parameters.
The new version takes a big step up: the Ryzen AI Max+ PRO 495 chip, up to 192GB of LPDDR5x unified memory, and support for 300-billion-parameter models and larger at 4-bit quantization. It is aimed squarely at AI developers, researchers, and creators who want datacenter-class local AI without renting cloud GPUs.
Ryzen AI Max+ PRO 495: Full Specifications
The heart of the platform is the Ryzen AI Max+ PRO 495, which combines three processors on a single chip sharing one memory pool:
- CPU: Up to 16 Zen 5 cores (32 threads), boosting up to 5.2GHz from a 3.1GHz base
- GPU: 40 AMD RDNA 3.5 compute units (Radeon 8065S integrated graphics)
- NPU: XDNA 2 neural processor rated up to 55 TOPS
- Memory: Up to 192GB LPDDR5x unified memory (up to 160GB allocatable as VRAM)
- Storage: 2TB SSD
- Wireless: Wi-Fi 7, Bluetooth 5.4
- FP16 AI compute: Up to 61 TFLOPS
AMD notes that the 55 TOPS NPU figure is a peak rating under optimal conditions, and real-world throughput varies with system configuration, model, and software version.
192GB of Unified Memory: Why It Matters for Local AI
The headline feature is the memory. In a normal PC, system RAM and graphics memory are separate pools, and large AI models have to squeeze into whichever is smaller. The Ryzen AI Max+ PRO 495 puts the CPU, GPU, and NPU on a single high-bandwidth memory pool — up to 192GB of it, with as much as 160GB assignable to the GPU as VRAM.
That changes what fits on a desk. AMD says the platform can hold 300-billion-parameter language models at 4-bit quantization entirely in memory, alongside datasets, applications, and working context. Developers can also run multiple large AI agents concurrently, each with its own KV cache — something that simply does not fit on systems with smaller GPU memory.
Built for AI Agents, Not Just Chatbots
AMD is positioning the platform around the shift from chatbots to AI agents — software that reads files, calls tools, coordinates parallel processes, and produces finished deliverables like decks, memos, and charts. In that setup, each part of the chip has a job: the CPU orchestrates the agent’s tools and processes, the GPU and NPU accelerate model inference and content generation, and the unified memory keeps models and working data loaded for all three.
AMD argues that raw tokens-per-second still matters, but for agent workloads, orchestration, memory capacity, and end-to-end task completion matter just as much.
Generative AI Performance: AMD’s Claims vs Intel and Apple
AMD tested the PRO 495 on native x86 Windows with ROCm software across 18 image, video, music, and 3D generation workloads — including FLUX.2, Qwen Image, Ideogram V4, and Stable Music — and published comparisons against two rivals:
- Vs Intel Core Ultra X9 388H: Higher relative throughput across all 18 workloads, with a top result of 32.2x.
- Vs Apple M5 Pro: Higher throughput across the workloads, with a top result of 34.3x. AMD notes several workloads did not run on the Apple system at all.
Two caveats worth knowing: these are AMD’s own tests from September 2026, not independent reviews, and the comparison systems carried 64GB of memory against the Halo’s 192GB — a gap that decides which workloads can even run on each machine.
Software: Ryzen AI Developer Center and AI Playbooks
Hardware is only half the story. AMD bundles a software layer meant to take developers from first boot to a running model with minimal setup:
- Ryzen AI Developer Center: A single app for software, platform controls, features, and updates, with remote access and software sync.
- Preloaded apps and models: Installed out of the box so testing can start immediately.
- AI Playbooks: Guided paths through supported AI workflows.
- Best-Known Configuration: AMD validates ROCm software, drivers, frameworks, applications, and models together as one set, restorable through the Developer Center.
Windows-based versions will ship with Microsoft’s Windows Project Zenith, a ready-to-code Windows experience with VS Code, WSL, GitHub Copilot CLI, and PowerShell preinstalled.
Price and Availability
The Ryzen AI Halo Developer Platform with the Ryzen AI Max+ PRO 495 is available to order now from Micro Center in the US at $7,299.99 — $3,300 more than the first-generation 128GB model. Note that Micro Center lists it as in-store pickup only, with no shipping. Windows Project Zenith versions are coming soon. AMD has not announced pricing for other regions.
Ryzen AI Halo vs NVIDIA RTX Spark: The Local AI PC Race
The timing is no coincidence. NVIDIA just opened pre-orders for its RTX Spark AI PCs — laptops and compact desktops with up to 128GB of unified memory starting at $2,599. AMD’s answer is a more expensive, more memory-heavy developer box: 192GB vs 128GB, $7,299 vs $2,599.
The two target different buyers. RTX Spark is a consumer-and-creator platform that also plays AAA games; the Ryzen AI Halo is a developer-first machine for people whose models genuinely need 160GB of VRAM. But together they mark the same turning point: 2026 is the year serious AI moved from the cloud to the desk.
Should You Buy It?
The Ryzen AI Halo makes sense for a narrow audience: AI developers and researchers running very large local models or multi-agent workflows who would otherwise pay for cloud GPU time. For everyone else — including most creators — the $7,299 price and developer focus make it a tough sell against cheaper local-AI options.
Still, as a statement of where the industry is going, it matters. Unified memory architectures are rapidly becoming the default for local AI, and AMD just raised the bar for how much of it you can get in a box that sits quietly on a desk.