That distinction matters. A late-August report can make the configuration seem newly imminent, yet substantially the same 192GB Ryzen AI Max+ PRO 495 configuration and headline specifications were reported in July. The current evidence shows an announced forthcoming option, not a newly confirmed launch milestone. For Windows users planning a local AI workstation, the sensible conclusion is to track it, not to treat it as hardware that can be bought or benchmarked today.
What Framework is proposing
The forthcoming top Framework Desktop configuration is listed with AMD’s Ryzen AI Max+ PRO 495 processor and 192GB of LPDDR5X unified memory. Its stated specifications include 273GB/s memory bandwidth, 131 TOPS of total AI compute, a 16-core, 32-thread Zen 5 CPU, and a Radeon 8065S integrated GPU with 40 compute units. A Linux preload is also presented as an option.
The unusually large unified-memory pool is the central story. Traditional desktop configurations divide system RAM and GPU VRAM into physically separate pools. Unified memory instead allows the CPU and integrated GPU to access the same overall memory resource. That can be especially attractive for local model inference, where fitting a model and its working data into available memory can be a more immediate obstacle than raw processor speed.
It is important, however, not to mistake unified capacity for dedicated-GPU equivalence. A 192GB memory figure describes the system’s total unified pool, not 192GB of graphics memory guaranteed to applications. The amount that firmware and software can expose or reserve for graphics workloads is a separate implementation detail, and Framework has not published that detail for this particular Desktop configuration.
What changes with Ryzen AI Max+ PRO 495
AMD’s published specifications show that the PRO 495 is a measured iteration over the prior Ryzen AI Max+ PRO 395, rather than a complete redesign. Both chips have 16 Zen 5 CPU cores and 32 threads, 40 graphics cores, and a configurable TDP range of 45W to 120W. That baseline suggests buyers should not expect a radically different class of CPU or integrated-GPU performance solely from the processor name.
The improvements are real, but specific:
- Maximum CPU boost rises from 5.1GHz to 5.2GHz.
- Maximum graphics frequency rises from 2.9GHz to 3.0GHz.
- NPU performance increases from 50 TOPS to 55 TOPS.
- Total quoted AI compute increases from 126 TOPS to 131 TOPS.
- Maximum supported memory grows from 128GB of LPDDR5X-8000 to 192GB of LPDDR5X-8533.
The last item is more consequential than the small frequency gains for the type of buyer likely considering this product. AMD specifies a 256-bit LPDDR5X-8533 interface for the PRO 495, compared with 256-bit LPDDR5X-8000 for the PRO 395. In theoretical terms, that works out to 273.056GB/s versus 256GB/s: about a 6.66% bandwidth increase.
That calculation aligns with Framework’s 273GB/s headline. Still, it is a theoretical interface figure, not a performance result. Actual gains depend on the application, memory access patterns, operating system, drivers, cooling, power limits, and the software backend used for an AI model. A memory-bound workload may benefit more than a CPU-bound task; another application may see little change.
The capacity claim needs careful reading
AMD says the PRO 495 platform supports up to 160GB of dedicated graphics memory and describes that capacity as capable of running models exceeding 300 billion parameters at 4-bit quantization. Those statements point to an important possibility for large local models, but they are processor-platform claims from AMD, not tested results for the finished Framework Desktop.
Several practical qualifications follow.
First, AMD’s 160GB dedicated-graphics-memory statement is not the same thing as Framework advertising 192GB of unified system memory. The former concerns a potential graphics allocation; the latter is a total system-memory specification. Without Framework documentation for its firmware configuration, buyers cannot assume precisely how much of the 192GB will be usable by a particular graphics or inference runtime.
Second, a parameter-count claim does not establish a universally usable experience. A model’s memory requirements are affected by quantization, context length, runtime overhead, model architecture, operating system memory use, and the software implementation. Fitting a model is not identical to running it at a useful speed.
Third, no reviewed material provides independently reproducible inference throughput for a completed PRO 495 Framework Desktop. There are also no like-for-like comparisons showing it replacing multiple high-end discrete GPUs for a named model, quantization level, context length, and software stack. That popular shorthand may be tempting because large unified memory is rare at this size, but it goes beyond the available evidence.
For Windows users, this means the system may eventually be highly interesting for experimentation with local language models, image-generation workflows, development work, and tasks that benefit from ample shared memory. It does not yet justify conclusions about tokens per second, gaming performance, professional rendering, thermal behavior, noise, or sustained performance under load.
A potentially useful expansion detail
One report on the planned 192GB model describes an open-ended PCIe x4 slot intended to accommodate physically larger PCIe cards. This could be meaningful for a small desktop aimed at advanced users. Depending on the final platform configuration and card support, such a slot could provide options for networking, storage, capture hardware, or other expansion beyond what a closed mini PC typically allows.
The same report says Framework demonstrated two systems fitted with 50GbE cards in a 384GB RoCE cluster. That demonstration illustrates a possible direction for networking and multi-machine experimentation, especially for users exploring distributed workloads. It is not, however, independent proof of the performance, availability, or standard configuration of the final 192GB Desktop.
Buyers should therefore treat the PCIe detail as promising but provisional. Important unanswered questions include whether the open-ended slot is present on every retail 192GB configuration, which lane generation and bandwidth it uses in the shipping system, what power and physical-clearance limits apply, and whether any demonstrated networking setup requires optional hardware.
Why a competing announcement is not a Framework timetable
AMD’s PRO 495 is beginning to appear in other vendors’ launch plans. GMKtec has scheduled a September 5, 2026 event for a product powered by the processor, with independent reporting identifying the system as the EVO-X5 Pro.
That confirms that the chip is moving into the wider market. It does not establish that Framework will announce, take orders for, or ship its own Desktop on the same date. Product schedules are set by the individual system maker and can be shaped by memory supply, validation, firmware, regional logistics, and support readiness. Readers should resist connecting the two timelines without an explicit Framework statement.
The buying decision: wait for data, not just specifications
The 192GB Framework Desktop has a coherent technical proposition. The PRO 495’s biggest advance over the PRO 395 is not a dramatic core-count change; it is the combination of higher supported memory capacity and slightly higher memory bandwidth. For local AI users whose projects repeatedly run into VRAM or system-memory ceilings, that may be more valuable than modest peak-clock improvements.
Yet the information still missing is unusually important because this will likely be a premium configuration. Pricing remains unannounced. Earlier reporting said the model was expected to represent a substantial step up from the 128GB Desktop’s price, but there is no specific, reliable price figure to use for budgeting. Buyers also lack a release window, regional order information, final configuration details, and third-party testing.
Windows-focused purchasers should wait for several concrete answers before committing:
- Final price and whether storage, operating system, and accessories materially change it.
- A preorder or shipping date rather than another “coming soon” listing.
- Windows driver support and real-world testing for the Radeon 8065S and relevant AI frameworks.
- Independent results for local inference, graphics workloads, power use, heat, and acoustics.
- Framework’s actual graphics-memory allocation options and documented software support.
- Full PCIe expansion specifications for the 192GB model.
There is also a broader trade-off. Unified-memory systems can make previously impractical large-model experiments possible in a compact form factor, but a discrete-GPU workstation may still be the better choice for users who need established CUDA-dependent tools, known acceleration behavior, easier GPU replacement, or verified performance in a specific professional application. Which direction is better depends less on headline TOPS than on the software a buyer must run.
The bottom line
Framework’s 192GB Ryzen AI Max+ PRO 495 Desktop is worth watching because it pairs exceptionally large LPDDR5X unified memory with a processor whose formal improvements are tailored toward that capacity. The 273GB/s bandwidth figure is consistent with AMD’s LPDDR5X-8533 specification, and the move from a 128GB to 192GB maximum is a meaningful platform change.
But the evidence does not show that Framework is closer to launch than it was when this configuration first became public. It remains a coming-soon product without a price, order date, delivery target, or independently verified performance profile. Until those details arrive, the 192GB Desktop is best understood as a potentially important local-AI workstation option—not yet a completed buying choice.