AMD's new mini PC runs 300-billion-parameter AI models on your desk
AMD just dropped a desktop rig with 192GB of unified memory, letting you run massive 300B parameter AI agents locally without the cloud.
Tools · Source: AMD
What happened
AMD announced the next Ryzen AI Halo Developer Platform. It is a compact desktop system built strictly for local AI workloads. The machine runs on the new AMD Ryzen AI Max+ PRO 495 processor. It packs up to 192GB of LPDDR5x unified memory. This hardware sits quietly on your desk but delivers massive capacity.
This unified memory pool completely changes how local AI works. The CPU, GPU, and NPU share a single high-bandwidth memory pool instead of splitting system RAM and graphics memory. You can allocate up to 160GB directly to the GPU as VRAM. This gives developers enough room to run 300-billion-parameter models at 4-bit quantization.
You can order the system right now from Micro Center. Windows-based versions equipped with Microsoft Windows Project Zenith are coming soon. AMD bundles a Developer Center application to get builders from first boot to first token in minutes. The platform includes preloaded apps, models, and guided AI playbooks.
Key facts
- 192GB — Maximum LPDDR5x unified memory available on the platform
- 300 billion — Parameter size of models supported at 4-bit quantization
- 160GB — Maximum memory that can be allocated to the GPU as VRAM
- 55 TOPS — Peak NPU performance rating under optimal conditions
- 16 — Maximum number of Zen 5 CPU cores on the processor
Why it matters
AI is shifting rapidly from simple chatbots to complex agents. Agents do more than generate text responses. They read files, call external tools, coordinate parallel processes, and produce finished deliverables. This requires massive memory to hold multiple models, datasets, and active working context at the exact same time. AMD is optimizing for end-to-end task completion instead of just measuring tokens per second.
Massive local memory means you stop relying on cloud providers for heavy workloads. You keep sensitive data entirely under your control. Hospitals, finance firms, and enterprise developers can run multi-agent workflows right on their desks. If your data never leaves the room, privacy compliance becomes a non-issue. You connect to the cloud only when the work calls for it, not because your hardware forces you to.
For builders
Massive local VRAM unlocks multi-agent workflows
You can allocate 160GB of memory directly to the GPU. This lets you run multiple large agents concurrently with discrete KV caches. Enterprise developers save money on cloud inference while building complex parallel systems. Startups building privacy-first tools win big here.
Out-of-the-box software environment saves time
AMD provides a validated Best-Known Configuration with ROCm software, drivers, and frameworks pre-installed. You spend less time troubleshooting setup errors and more time actually building products. Developers win by getting to their first token in minutes instead of days.
Hardware designed for end-to-end completion
Pure inference speed is no longer the only metric that matters. The CPU orchestrates tools, while the GPU and NPU accelerate generation. Builders creating complex deliverables get a balanced system that finishes the job. You stop worrying about memory bottlenecks during multi-step tasks.
My take
I have been waiting for hardware that actually understands where AI is going. Agents need memory more than they need raw speed, and 160GB of local VRAM is an absolute game changer for privacy. If you are building AI for finance or healthcare, keeping your data on your desk instead of sending it to OpenAI is your biggest competitive advantage. Stop renting cloud GPUs and buy your own metal.
Original reporting: AMD. This is my rewrite and opinion.