AMD Ryzen AI Max+ PRO 495: 192GB Unified Memory, Runs 300B Models Locally

<div className="tldr-box"> <h3>Key Takeaways</h3> <ul> <li>TL;DR AMD shook up the local AI hardware landscape with the Ryzen AI Max+ PRO 495, a monolithic APU delivering up to 192 GB of unified LPDDR5X8533 memory — 160 GB of which is allocatable to the integrated GPU.</li> <li>In a mini PC starting at roughly $3,999, this single chip can run 300Bparameter models at Q4KM quantization entirely offline.</li> <li>Early benchmarks show it outperforming the NVIDIA DGX Spark on midsize models while costing less and drawing a fraction of the power.</li> <li>Why This Matters The biggest bottleneck for running large language models locally has never been compute — it's memory bandwidth and capacity.</li> </ul> </div>

TL;DR

AMD shook up the local AI hardware landscape with the Ryzen AI Max+ PRO 495, a monolithic APU delivering up to 192 GB of unified LPDDR5X-8533 memory — 160 GB of which is allocatable to the integrated GPU. In a mini PC starting at roughly $3,999, this single chip can run 300B-parameter models at Q4_K_M quantization entirely offline. Early benchmarks show it outperforming the NVIDIA DGX Spark on mid-size models while costing less and drawing a fraction of the power.

Why This Matters

The biggest bottleneck for running large language models locally has never been compute — it's memory bandwidth and capacity. Even high-end consumer GPUs top out at 24 GB (RTX 4090) or 48 GB (RTX 6000 Ada). The Ryzen AI Max+ PRO 495 sidesteps the GPU VRAM wall entirely by using a unified memory architecture where the CPU and GPU share a single pool of up to 192 GB. This means:

• No model sharding for most open-weight models
• Full Q4 inference for 300B-parameter models
• Zero data transfer between CPU and GPU memory
• Silent, power-efficient operation at 45–120 W TDP

For enterprises, researchers, and privacy-conscious users, this is the first truly capable single-box local LLM workstation that doesn't require a server rack or a dedicated cooling solution.

Specs

| Specification | Detail | |---|---| | Memory | Up to 192 GB LPDDR5X-8533 unified | | GPU-allocatable memory | Up to 160 GB | | Memory bandwidth | ~546 GB/s (estimated) | | TDP | 45–120 W configurable | | Form factor | Mini PC (e.g., ASUS NUC 14 Pro+ variant) | | Price | ~$3,999 (complete mini PC system) | | CPU cores | 16 Zen 5 cores / 32 threads | | GPU compute units | 40 RDNA 3.5 CUs | | AI accelerators | 2x XDNA 2 NPU (50 TOPS total) |

What You Can Run

At Q4_K_M quantization, here's what fits entirely in the 160 GB GPU-addressable pool:

| Model | Size (Q4_K_M) | Fits? | Notes | |---|---|---|---| | Llama 3.3 70B | ~38 GB | Yes | Plenty of headroom for context | | Qwen 2.5 72B | ~40 GB | Yes | Full 128K context possible | | DeepSeek V3 67B | ~37 GB | Yes | | | Mixtral 8x22B | ~52 GB | Yes | | | Command R+ 104B | ~58 GB | Yes | | | 300B models | ~165 GB | Yes | Tight but fits at Q4_K_M | | Llama 4 400B (MoE) | ~110 GB | Yes | MoE sparse compute is ideal |

Benchmarks

AMD's internal benchmarks (May 2026) show:

• GLM 4.7 (30B): 14% faster token throughput than the NVIDIA DGX Spark ($4,679 system)
• Llama 3.3 70B (Q4): ~7–9 tok/s at 4K context
• 300B model (Q4_K_M): ~1.5–2.5 tok/s — usable for batch inference and background processing
• Power draw during inference: 65–85 W versus 150–250 W for a discrete GPU workstation

Source: AMD Internal Benchmarks, May 2026

Performance scales with power budget — setting TDP to 120 W yields approximately 30% higher throughput than 45 W.

ROI vs. Cloud

| Factor | AMD Ryzen AI Max+ PRO 495 | Cloud (AWS p4d.24xlarge) | |---|---|---| | Upfront cost | $3,999 | $0 | | Monthly cost | ~$20 (electricity) | ~$32,000 (on-demand) | | 3-year TCO | ~$4,700 | ~$1,152,000 | | Data privacy | Full | None | | Latency | Local (no network) | Variable | | Scalability | Single user/small team | Elastic |

The payback period versus cloud inference is roughly 4 days if you would otherwise run a high-end GPU instance for 8 hours daily.

Software Compatibility

| Software | Status | |---|---| | Ollama | Fully supported (llama.cpp backend) | | LM Studio | Supported | | llama.cpp | Native support via Vulkan/BLAS | | Open WebUI | Works out of box | | ComfyUI | Supported (via ROCm fork) | | PyTorch / TensorFlow | ROCm 6.x support | | vLLM | Experimental (CPU-offload mode) |

Where to Buy

• ASUS — NUC 14 Pro+ with Ryzen AI Max+ PRO 495 (Q3 2026)
• Minisforum — MS-A1 Pro variant (pre-order, Q4 2026)
• Simply NUC — Custom configuration (available now)
• ThinkMate — RM Series silent workstation (Q4 2026)

FAQ

Can I upgrade the memory later?

No. The 192 GB LPDDR5X is soldered as part of the package. Choose your capacity at purchase time.

Will this run CUDA-dependent AI software?

Most CUDA code needs to run through ROCm or Vulkan translation layers. AMD has published a compatibility layer, but some niche tools may not work. Check the ROCm compatibility matrix.

How does it compare to a Mac Studio with 192 GB?

The AMD APU has higher memory bandwidth (546 GB/s estimated vs. 400 GB/s on M2 Ultra) and wider software compatibility for non-Apple ecosystems. The Mac Studio's GPU compute is faster for fp16 workloads, but the AMD wins on price-per-GB-of-memory.

Is 120 W enough to cool silently?

Yes — the mini PC form factor uses a large vapor chamber cooler. At 65 W inference loads, the system is virtually silent. At 120 W sustained, fan noise is comparable to a 14-inch MacBook Pro under load.