Framework Sets September 30 Preorders for 192GB Ryzen AI Max+ PRO 495 Desktop

Framework says preorders for the 192GB version of its compact Desktop will open Wednesday, September 30, in the morning Pacific Time. The system was previewed earlier this year around AMD's Ryzen AI Max+ PRO 495 platform, but the new announcement turns the previously 'coming soon' configuration into a dated preorder milestone.

The headline specification is 192GB of unified LPDDR5X memory. Framework's existing Ryzen AI Max+ 395 Desktop tops out at 128GB, so the new configuration increases capacity by 50%. That extra capacity is aimed primarily at local AI and workstation workloads where model size can be constrained by available GPU-addressable memory rather than raw compute alone.

The move from LPDDR5X-8000 to LPDDR5X-8533 raises theoretical memory bandwidth from 256 GB/s to 273 GB/s, an increase of about 6.6%. That distinction matters for local AI: 192GB lets larger models fit in unified memory, but it should not be interpreted as a 50% inference-speed increase. Workloads limited by memory bandwidth or compute can scale very differently from workloads limited by capacity.

Framework has demonstrated the system running DeepSeek-V4-Flash at Q8 and says up to 160GB of the unified pool can be allocated to the GPU. Those are useful capability indicators, not independent Core Tech Tips performance measurements. Framework has not supplied a general tokens-per-second result that can be applied across models, quantizations and software stacks.

The new configuration uses an open-ended PCIe x4 slot, allowing physically longer cards to fit while remaining electrically x4. Framework has shown two systems fitted with 50GbE adapters and connected using RDMA over Ethernet as a demonstration of multi-system local-AI work. That does not turn the two machines into one physically shared memory pool; distributed software still has to partition and communicate the workload across the systems.

Framework also says the 192GB version will include a pre-installed Noctua fan and that a pre-built Fedora configuration will be available. Those details reinforce the workstation and local-AI positioning rather than making the new model a conventional discrete-GPU gaming tower.

Unified-memory APUs occupy a different niche from PCs built around discrete gaming GPUs. A high-end discrete GPU can provide substantially more dedicated graphics bandwidth, but its VRAM capacity is separate from system RAM. A 192GB unified pool gives software a much larger addressable capacity for models that can use the architecture effectively.

The tradeoff is that capacity, bandwidth and software support must be evaluated separately. The September 30 preorder opening will finally add price and ordering information to that equation; independent workload testing will still be needed to establish how the PRO 495 configuration performs against discrete-GPU workstations and other high-memory local-AI systems.