AMD has introduced the next Ryzen AI Halo developer platform, a compact desktop system built around the Ryzen AI Max+ PRO 495, with up to 192GB of unified memory and up to 160GB of that assignable to the GPU as VRAM, according to AMD’s launch post. AMD puts the ceiling at models of 300 billion parameters and larger at 4-bit quantization, up from 200 billion on the first Halo, which AMD announced in May with a Ryzen AI Max+ 395 and 128GB. Orders are open through Micro Center.
The extra 64GB of memory is what moves the ceiling from 200 to 300 billion parameters
The processor has 16 Zen 5 cores, 40 RDNA 3.5 compute units and an NPU rated up to 55 TOPS, all sharing one pool of LPDDR5x. The May announcement lists the 495 at up to 5.2 GHz boost, 80 MB of cache and a 45 to 120W configurable power range, so the generational change is mostly capacity: 192GB against 128GB, and a 160GB graphics allocation. A Windows-based version running the new Windows Project Zenith is listed as coming soon.
The model limit follows from simple arithmetic on that allocation. Four bits per parameter makes 300 billion parameters about 150GB of weights, which leaves roughly 10GB of the 160GB graphics limit, and 32GB of the 192GB total for the operating system and everything else. The first Halo’s 200 billion parameters took about 100GB of its 128GB. AMD also describes running several large agents at once, each with its own KV cache. Those caches share whatever the weights leave over, so about 10GB at the 300-billion-parameter limit, and more as the model shrinks. This ignores quantization overhead such as scales, which adds to the weights.
AMD’s image and video comparisons set a 192GB system against two 64GB machines
For performance, AMD shows relative throughput in ComfyUI 0.37, tested in September, against an Intel Core Ultra X9 388H system with Arc B390 graphics and against an Apple M5 Pro MacBook Pro. AMD’s test system had 192GB of memory with 128GB set aside for graphics and ran Adrenalin 26.8.1; the Intel and Apple machines each had 64GB. AMD reports higher throughput for the 495 across the workloads it ran, and several of those workloads did not run on the Apple machine, so no ratio was calculated for them.
The workloads are generative image, video, music and 3D tools, among them FLUX.2, Qwen Image, LTX, ACE-Step and Microsoft Trellis. The text of the post gives no tokens-per-second figure for the large language models that the 192GB headline is about; the post says instead that orchestration, memory and end-to-end completion matter as much as model inference.
The post lists no price. It points buyers to Micro Center, the same exclusive retailer named for the first Halo.