Framework Desktop with Ryzen AI Max+ PRO 495: a 192GB RAM mini PC aimed at on-device AI

Framework has previewed a new version of its Framework Desktop built around the AMD Ryzen AI Max+ PRO 495 processor, a 16-core Zen 5 chip clocked at 5.2 GHz paired with the Radeon 8065S integrated GPU. The headline feature is a rare 192GB of LPDDR5x unified memory, enough to run large models like DeepSeek-V4-Flash at Q8 quantization locally with room for context length. This is a clear signal that on-device AI hardware is moving beyond thin clients into serious workstation territory.

What happened

At AMD's Advancing AI event, Framework showed off a desktop that combines the Ryzen AI Max+ PRO 495 with 192GB of RAM in a compact 4.5-liter chassis. The system also includes an open-ended PCIe x4 slot for adding a dedicated GPU, and it ships with either Windows or a Linux pre-install option. Framework also demonstrated a proof of concept where two desktops were linked via 50GbE networking to pool 384GB of memory across both machines.

Pricing context is important. The current 128GB version with the Ryzen AI Max+ 395 costs $3,449, up from $1,999 at launch in February 2025 due to rising RAM and storage costs. Framework has warned that the 192GB model will be a substantial jump, and analysts expect it could exceed $4,500. Official pricing and a release date have not been announced.

Why AI builders should care

For developers and product teams building AI applications, the Framework Desktop represents a shift toward local inference that reduces reliance on cloud APIs. Running models on-device eliminates per-token API costs, lowers latency, and keeps data local for privacy-sensitive workloads. The 192GB unified memory pool is particularly relevant for hosting large models that previously required multi-GPU cloud instances.

The Ryzen AI Max platform also supports AMD's AI software stack, which could simplify deployment for teams already using ROCm or ONNX Runtime. The ability to allocate a large portion of the 192GB as VRAM means you can run models that would otherwise need a dedicated GPU with expensive HBM memory.

Practical implications

If the pricing lands within reach, the Framework Desktop could become a viable option for several use cases:

• Local model experimentation: Run quantized versions of 70B+ parameter models without cloud costs.
• CI/CD for AI pipelines: Test model changes on a local machine before deploying to production.
• Edge deployment prototyping: Validate inference performance on hardware that resembles edge devices.

The PCIe x4 slot adds flexibility for attaching an external GPU or other accelerators, and the 50GbE networking proof of concept hints at clustered setups for larger models. The dual-OS support (Windows or Linux) means teams can use their preferred development environment.

Caveats

Several important details remain unknown. Framework has not disclosed official pricing or a release date for the 192GB model. The claimed capability to run DeepSeek-V4-Flash is based on a preview and proof-of-concept demonstration, so real-world performance may vary with firmware and software optimizations. The 256 GB/s memory bandwidth of the LPDDR5x could also become a bottleneck for inference on very large models, as some community analysis has noted that at this bandwidth, token generation rates may be limited.

Finally, the price trajectory is concerning. The 128GB model has already risen 72% since launch, and the 192GB version could push past $4,500. At that price point, the value proposition versus cloud inference or a custom-built workstation needs careful evaluation.