Question about Framework Desktop 192 GB as a local AI workstation (Strix Halo + dGPU)

Framework Desktop
, , ,
1

Hi everyone,

I’m very interested in the upcoming Framework Desktop 192 GB configuration, especially for running large local language models such as DeepSeek V4 Flash.

I’ve been following the recent work around DwarfStar, Lucebox and other runtimes optimized for Strix Halo, and the new 192 GB SKU looks like it could become a very compelling platform for local AI.

Before placing an order, I have a few questions that I hope someone from Framework (or anyone familiar with the hardware) might be able to answer.

PCIe slot

Will the 192 GB version keep the same PCIe implementation as the current model?

Specifically:

  • Is the slot still electrically PCIe 4.0 x4?

  • Are there any changes planned for lane allocation or firmware?

  • Is there any reason why a discrete GPU connected through a high-quality PCIe Gen4 x4→x16 riser would not enumerate correctly?

I understand the bandwidth limitations of x4 and I’m not expecting gaming performance. My interest is purely AI inference.

Discrete GPU for AI

One configuration I’m considering is:

  • Framework Desktop 192 GB

  • Radeon 7900 XTX or similar (either through x4 connector and a riser or through a x16 connector if available)

  • External ATX power supply for the GPU

The goal would be to experiment with asymmetric execution (similar to what Lucebox has recently demonstrated), where the Strix Halo iGPU and unified memory hold most of the model while the discrete GPU accelerates the dense path and the hottest experts.

Has anyone from Framework tested this kind of setup?

AI-focused use case

The 192 GB model seems particularly attractive because it should comfortably accommodate:

  • DeepSeek V4 Flash

  • large KV caches

  • draft/speculative models

  • additional runtime overhead

Has anyone internally experimented with this type of workload?

I’m not asking for official support for Lucebox or DwarfStar—I’m simply curious whether Framework sees local AI as one of the intended use cases for this platform.

I’d really appreciate any technical information you can share.

Thanks!

2

In case it’s not clear, this device is not available for purchase yet. I would love to hear that it has gone into production, but I don’t think that has been confirmed.

I’d love to get one myself, but I found the 128GB cost-prohibitive, and the price of the 192GB is going to be bonkers. I’ve got an AI use-case myself, but currently I’m not sure that local LLMs can do code-gen at the quality I get from cloud providers.

Some price speculation here: FW Desktop with 192GB RAM & AI MAX+ PRO 495 coming soon

1 Like
3

In relation to your questions, I am all ears on them myself. :smiley:

However FW is always tight-lipped about specs/progress until they release the same in formal marketing channels, so you might have to just put up with this teaser.

4

Yes, very much :slight_smile: We had a local LLM demo at AMD Advancing AI event a few weeks ago.

You can find some information about the new desktop here : https://frame.work/blog/framework-laptop-13-pro-press-reviews-and-previewing-a-192gb-framework-desktop