Can I Run Devstral on Framework Desktop (Ryzen AI Max+ 395, 128 GB)?

Autor: Jakub Rusinowski · Ostatnia aktualizacja: 11 września 2026

Technically yes, but not recommended

It loads, but it is not worth running — Devstral-2 123B at Q6_K fits in Framework Desktop (Ryzen AI Max+ 395, 128 GB)'s 128 GB, yet the memory bandwidth limits it to ~1.9 tok/s (estimated), well below usable interactive speed.

Confidence: medium · Recommended quantization: Q6_K · Estimated speed: ~1.9 tok/s

Get a personalized upgrade path →

Uruchom w chmurze RunPod

lub porównaj w Vast.ai

Jako partner Amazon zarabiamy na kwalifikujących się zakupach. Linki do chmurowych GPU są linkami polecającymi — możemy otrzymać prowizję bez dodatkowych kosztów dla Ciebie.

Framework Desktop (Ryzen AI Max+ 395, 128 GB) — what it gives a model

Usable memory for models	128 GB
Memory bandwidth	256 GB/s
Form factor	Mini PC
Operating system	Windows or Linux
Memory upgradeable	No — soldered

Devstral on Framework Desktop (Ryzen AI Max+ 395, 128 GB): memory by quantization

Quant	Memory needed	Fits 128 GB?	Max context	Est. speed	Download
F16	249.7 GB	✗ No	—	—	246 GB
Q8_0	134.4 GB	✗ No	—	—	130.7 GB
Q6_K	104.5 GB	✓ Yes	64K	~1.9 tok/s	100.9 GB
Q5_K_M	90.9 GB	✓ Yes	64K	~2.2 tok/s	87.2 GB
Q4_K_M	77.9 GB	✓ Yes	128K	~2.5 tok/s	74.3 GB
Q3_K_M	56.1 GB	✓ Yes	128K	~3.5 tok/s	52.4 GB
Q2_K	44.1 GB	✓ Yes	128K	~4.5 tok/s	40.4 GB

Which Devstral sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Devstral-2 123B	77.9 GB	✓ Fits	~2.5 tok/s
Devstral Small 2505 24B	16.6 GB	✓ Fits	~12.3 tok/s
Devstral Small 2 24B	16.6 GB	✓ Fits	~12.3 tok/s

What to watch out for

At ~1.9 tok/s this loads but is too slow for interactive use — expect roughly 32 seconds per 60 tokens.
This model does not publish its full attention configuration, so the KV-cache share of these figures is inferred from its parameter count rather than computed exactly.
Memory on this machine is not upgradeable, so the configuration you buy is the ceiling for every model you will ever run on it.

Framework Desktop 128 GB limitations

Memory is soldered LPDDR5X — unusually for Framework, this is the one component that cannot be upgraded.
ROCm/Vulkan support for Strix Halo is younger than CUDA; check your runtime supports it before buying.

Recommended setup

Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
128 GB unified memory at 256 GB/s, shared between CPU and GPU.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is inferred from the parameter count because this model does not publish its full attention configuration.

FAQ