Framework Desktop (Ryzen AI Max+ 395, 128GB) with LM Studio: official getting-started guide
Framework's own guide to running local LLMs on Framework Desktop with LM Studio: full GPU offload, Mistral Small 3.2 24B at ~10 tok/s and gpt-oss-120b at ~40 tok/s on the 128GB model (faster on Linux).
Solo2-10 peopleWriting assistantCodingDocument search / chat with files$3k+Sensitivity: High (client/confidential)
Hardware
| Item | Approx. price (as of 2026-10-16) |
|---|---|
| On sale −20% Minisforum MS-S1 MAX (Ryzen AI Max+ 395, 128GB, 2TB) (Minisforum Store, related current offer) | $3,879.00 |
| On sale −6% MSI GeForce RTX 5060 Ti 16GB Ventus 3X OC (Newegg, related current offer) | $579.99 |
| Framework Desktop, Ryzen AI Max+ 395, 128GB (32GB/64GB also discussed) | Not stated in source |
| Framework Desktop DIY (Ryzen AI Max+ 395, 128GB) (Framework, related current price) | $3,449.00 |
Source prices are as stated by the source. Retailer offers: prices checked 2026-10-16, may change. Some links may be affiliate links; we may earn a commission at no extra cost to you. This never changes what we list.
Software stack
- LM Studio (llama.cpp)
- Windows 11 or Fedora 43
Models used
- Mistral Small 3.2 24B (Q6_K)
- gpt-oss-120b
Good at
- Up to 112GB addressable by the GPU
- gpt-oss-120b 'around 40 tok/s (48 tok/s on Linux)'
- Inference about 20% faster on Fedora 43 than Windows 11 (as of Dec 2025)
Weak at
- Dense 24B model only 'around 10 tok/s (12 tok/s on Linux)'
- 256GB/s memory bandwidth is the typical bottleneck
Rough total cost
Not stated in source.
Getting started
- Install the Framework Desktop driver bundle and LM Studio.
- Download a model (e.g. Mistral Small 3.2 Q6_K or gpt-oss-120b on 128GB).
- Set GPU Offload to maximum and load the model.
- Optionally raise context length with Flash Attention enabled.
Source & credit
Using a Framework Desktop for local AI (Nirav Patel, updated Dec 2025) — Framework blog
Summarized from the source; quotes are the source author's words. Check the original before buying.